Training-Free Cultural Alignment of Large Language Models via Persona Disagreement
Dao Sy Duy Minh1,† [email protected]
Huynh Trung Kiet1,† [email protected]
arXiv:2605.10843v1 [cs.CL] 11 May 2026
Tuan Nguyen2,† [email protected] Chi-Nguyen Tran1 [email protected]
Phu-Hoa Pham1 [email protected]
Nguyen Lam Phu Quy1 [email protected] The Anh Han3 [email protected]
Long Tran-Thanh2,∗ [email protected]
Abstract LLMs are increasingly deployed in decisions that require culture-dependent moral judgements, yet they answer as if the whole world thinks with a Western mindset. The Moral Machine experiment [Awad et al., 2018] showed this is wrong at scale: 40 million judgments across 233 countries reveal that moral preferences are systematically structured by culture, and a model that ignores this variation does not merely underperform, but also imposes one society’s intuitions on all others. Existing fixes do not scale to global deployment, as fine-tuning needs per-country preference data and GPU budgets, reward-guided decoding needs per-country reward models, and activation steering needs access to model internals that blackbox APIs do not expose. In this work, we focus on this realistic inference time regime, with no weight updates, no training data, and no internal access. The key observation is that within-country demographic disagreement, not consensus, is the steering signal. When culturally grounded personas agree, the base model is already calibrated. But when they disagree, the spread tells us what to fix and how. We propose DISCA, which instantiates each country as a panel of four WorldValues-Survey-grounded persona agents [Haerpfer et al., 2020]. It converts their disagreement into a bounded, loss-averse correction, whose magnitude is set by the panel’s variance, and shrinks the correction toward zero when the estimate is unreliable. Across 20 countries and 7 open-weight backbones from five model families, DISCA reduces cultural misalignment on MultiTP [Jin et al., 2025] by 10–24% on the six ≥3.8B backbones, 3.4% on the smallest 2B model on binary moral dilemmas, and 2–7% on open-ended scenarios. Furthermore, a 14B backbone with DISCA reaches lower absolute misalignment than a vanilla 70B model. 1 Faculty of Information and Technology, University of Science, Vietnam National University, Ho Chi Minh City, Vietnam. 2 Department of Computer Science, University of Warwick, Coventry, United Kingdom. 3 School of Computing, Engineering and Digital Technologies, Teesside University, United Kingdom. † Equal contribution. ∗ Lead senior author.
Preprint.
1
Introduction
Large language models increasingly mediate decisions that turn on moral judgement: content moderation policies, clinical prioritisation guidelines, autonomous-vehicle behaviour specifications, and high-stakes recommendation pipelines. In each of these settings, the model’s implicit moral preferences become the system’s moral preferences. Yet a growing body of evidence shows that these implicit preferences are not culturally neutral. The Moral Machine experiment collected 40 million judgments across 233 countries and showed that moral preferences are systematically structured by culture [Awad et al., 2018]; large language model (LLM) outputs correlate most strongly with respondents from Western, educated, industrialised, rich, and democratic populations [Henrich et al., 2010, Santurkar et al., 2023]; and post-training alignment introduces further unintended cultural shifts [Ryan et al., 2024, Zewail et al., 2026]. A model that ignores this variation does not merely underperform on a benchmark, but also imposes one society’s intuitions on every user it serves. The cultural gap is now well-documented at the model level. The MultiTP benchmark [Jin et al., 2025] extends the Moral Machine protocol to 107 languages with country-specific Average Marginal Component Effects (AMCEs), and reports consistently weak country-level alignment across all 19 evaluated LLMs. The asymmetries are substantial, e.g., Japan ranks first among countries on preference for sparing pedestrians, while China ranks 116th, and current models do not capture either extreme. Concurrent work shows that sociodemographic persona prompts shift LLM moral decisions two to four times more than they shift human decisions [Kim et al., 2025], indicating that the models are not insensitive to cultural framing, but that the framing they receive at deployment time does not match the cultural distribution they are asked to serve. Closing this gap is harder than it looks. Three families of methods have been proposed, and each demands resources that do not scale to the long tail of countries served by a single deployed model. Fine-tuning and culture-aware adapters require per-country preference datasets and GPU budgets for every target population [Zhang et al., 2026, Yao et al., 2025]. Reward-guided decoding requires a separate trained reward model per country [Khanov et al., 2024, Mudgal et al., 2024], and activation steering requires write-access to model internals [Arditi et al., 2025, Turner et al., 2024] that blackbox APIs do not expose. Any deployable solution must therefore work at inference time, must not assume per-country reward models or labelled preference data, and must not assume write access to model internals beyond what an API exposes. We refer to this as the black-box, public-data-only regime: the method we propose requires only API-level access to decision-token log-probabilities (exposed by every open-weight backbone we evaluate and by major commercial APIs that return logprobs), and we argue this is the only regime where cultural alignment can scale from a handful of curated countries to the hundreds a globally-deployed model actually serves. We observe that when certain culturally grounded persona prompts representing the same country face the same moral dilemma, the dispersion of their decoding-time logit gaps is itself informative. If they agree, the base model is already calibrated for that country and no correction is needed. But if they disagree, the spread of their gaps encodes both the direction and the magnitude of the needed correction. Our method, DISCA (Disagreement-Informed Steering for Cultural Alignment), instantiates each country as a panel of four persona agents grounded in World Values Survey microdata [Haerpfer et al., 2020], evaluates them in a single batched forward pass, and converts the panel’s variance into a bounded, loss-averse logit correction whose magnitude is automatically attenuated when the underlying estimate is unreliable. We show that this attenuation is principled: the underlying oracle correction is the MSE-optimal scalar shrinkage of the consensus, whose weight depends only on the within-panel variance (Proposition 1), and we approximate it with an empirical variance-aware shrinkage heuristic. Across 20 countries spanning four continents and seven open-weight backbones (2B–70B parameters) from five model families, DISCA reduces cultural misalignment on MultiTP by 10–24% on the six ≥3.8B backbones (and 3.4% on the smallest 2B model) on binary moral dilemmas, and 2– 7% on open-ended scenarios. The strongest absolute alignment is achieved by Phi-4 (14B), which surpasses a vanilla Llama-3.3-70B backbone despite using one fifth of the parameters; this suggests that, in the regime we study, calibration competes with scale rather than merely complementing it. An evaluation on the BLEnD factual cultural QA benchmark [Myung et al., 2024] reveals a clean scope boundary: DISCA steers values, where the decision reduces to a scalar logit gap, but does not transfer to factual recall, where the decision is a single token in a large vocabulary. Failure modes are architecturally rather than culturally determined and are diagnosable from the base model’s vanilla logit conditioning alone. Primary metrics (MIS and diagnostics) are defined in §4. Our contributions are summarised as follows: 2
• An inference-time cultural alignment method that requires no weight updates, no percountry reward models, and no white-box access. • A formal characterisation of disagreement-driven shrinkage. DISCA is the first inferencetime alignment method to use independent-run disagreement as a reliability signal. • An empirical evaluation across 20 countries, four continents, seven open-weight backbones, and three evaluation formats (binary moral dilemmas, open-ended scenarios, factual QA).
2
Related Work
Benchmarks and cultural evaluation. The Moral Machine experiment [Awad et al., 2018] established that moral preferences are culturally structured, and MultiTP [Jin et al., 2025] extends it to 107 languages with country-specific AMCEs; LLM alignment to these AMCEs varies widely across model families [Takemoto, 2024, Ahmad and Takemoto, 2025] and correlates most with WEIRD respondents [Santurkar et al., 2023, Henrich et al., 2010], with RLHF adding unintended cultural shifts [Ryan et al., 2024, Zewail et al., 2026]. Kim et al. [2025] further show that sociodemographic personas shift LLM moral decisions 2–4× more than they shift humans (a “partisan sorting” phenomenon absent in human respondents). This last finding directly motivates our design: rather than using personas to steer toward a single viewpoint, DISCA uses them to measure within-country disagreement and converts that into a bounded correction. Inference-time alignment. The pluralistic alignment vision [Sorensen et al., 2024] argues that inference-time methods are a natural vehicle for cultural adaptation, but existing approaches each sacrifice one of the three constraints we target: activation steering [Arditi et al., 2025, Turner et al., 2024] requires white-box access; reward-guided decoding [Khanov et al., 2024, Mudgal et al., 2024] needs per-country reward models; culture-aware adapters [Zhang et al., 2026, Yao et al., 2025] need per-culture training. Training-time prospect-theoretic alignment [Ethayarajh et al., 2024] uses the same PT value shape we adopt, but at training time; DISCA applies it at test time, per scenario, aggregating heterogeneous personas rather than shaping a trained policy. Distinction from superficially similar paradigms. DISCA may resemble ensemble methods architecturally, but not algorithmically. In self-consistency [Wang et al., 2023] and LLM debate [Du et al., 2024, Liang et al., 2024], diversity is noise to be eliminated via majority vote; standard ensembles average outputs to reduce prediction variance. In DISCA, diversity is the signal: the spread of persona logit gaps, not the mode, sets the correction magnitude. We show (Proposition 1) that the MSE-optimal scalar shrinkage of the consensus correction is a closed-form function of within-panel variance, making within-panel variance a sufficient statistic for correction reliability. Standard calibration (temperature scaling, MC-Dropout [Gal and Ghahramani, 2016, Kwon et al., 2026]) adjusts confidence isotropically; DISCA applies anisotropic, loss-averse corrections. DISCA is the first training-free method targeting country-level moral calibration; the method is grounded in controlas-inference, Prospect Theory, and importance-sampling variance control.
3
Disagreement-Informed Steering for Cultural Alignment
DISCA aligns the model by creating an adjustment derived from the agents’ consensus, and the agreement among the agents controls how much of it actually applies. The magnitude of the adjustment is not fixed: when the agents agree on what the preference, we apply the adjustment in full; when they disagree, we apply only a fraction of it, scaled by how much the agents disagree. This inverts the standard combination structure used by self-consistency [Wang et al., 2023], multi-agent debate [Du et al., 2024, Liang et al., 2024], and ensembles, where N candidate outputs are aggregated by majority or averaging and disagreement among them is treated as noise to be suppressed. DISCA instead treats disagreement as the quantity that controls how much adjustment to apply. 3.1
Setup
We consider a frozen language model fθ on a forced-choice task: each input x asks the model to choose between two options via decision tokens A and B. Let z(x) = [za , zb ] denote the model’s logits at the decision-token position, and let δ(x) = zb − za be the resulting decision gap. For target country c, we assume access only to fθ as black-box and a public characterisation of country c’s population (in our case, World Values Survey microdata [Haerpfer et al., 2020]); we do not assume per-country preference labels, reward models, or write access to θ. Our goal is to compute, from fθ 3
Figure 1: DISCA overview. Stage 1 builds WVS-grounded persona prompts for a trolley scenario in country c; Stage 2 runs a frozen large language model (LLM) on the base prompt and each persona, aggregates persona-level signals in logit space, and applies Prospect-Theory importance sampling (PT–IS) together with a dual-pass reliability gate to obtain the final sparing probability. Pseudocode and the six MultiTP attribute–temperature pairs provided in App. A1. alone, a per-scenario logit adjustment δ ⋆ (x, c) such that the adjusted gap δ(x) + δ ⋆ (x, c) produces decisions better aligned with country c’s preferences. Alignment is measured against a country-level reference vector; the benchmark and metric are described in §4. The remainder of Section 3 focuses on how δ ⋆ should depend on the disagreement, what formal property this dependence has, and how to estimate the correction. 3.2
The disagreement principle
Suppose we evaluate the same scenario under N culturally grounded prompts representing the same target country, each producing a logit gap δi . Each δi is a noisy estimate of what country c’s population-level preference would imply for this scenario. If the gaps cluster tightly, the N estimates converge on the same answer and the consensus is a reliable signal. If they are dispersed, the consensus is averaging over disagreeing estimates and is unreliable as a steering target. The principle is that within-agent dispersion, not consensus magnitude, controls how much of the consensus adjustment we should actually apply. Let δ1 , . . . , δN be the logit gaps produced by N agents on a single scenario, and let δh denote the unobserved logit gap consistent with country c’s population-level preference on that scenario. We model the agents as δi = δh +ηi , E[ηi ] = 0, Var(ηi ) = τ 2 , with the ηi assumed independent across agents and identically distributed with common variance τ 2 . This is a standard hierarchical setup: δh is a fixed-but-unknown population parameter, and each agent is a noisy view of it. The i.i.d. structure is a modellingP simplification, since all agents share base modelPfθ and the scenario. Define the consensus δ̄ = N1 i δi , the within-agent variance D2 = N 1−1 i (δi − δ̄)2 , the consensus adjustment ∆ = δ̄ − δbase , and the oracle adjustment ∆h = δh − δbase , where δbase is the logit gap produced by the unconditioned base prompt (without persona or country-related prompts) on the ˆ same scenario. We will apply a scaled version of ∆ as our adjustment: ∆(γ) = γ · ∆ for some γ ∈ [0, 1]. We choose γ by minimising the mean squared error against the unobserved oracle: 2 ˆ MSE(γ) = E ∆(γ) − ∆h . Proposition 1 (Variance-aware shrinkage - Proof in App. A2). Under the agent model above: (i) D2 is unbiased for τ 2 , and the variance of the consensus adjustment is Var(∆) = τ 2 /N . ˆ (ii) Among shrinkage estimators of the form ∆(γ) = γ · ∆ with γ ∈ [0, 1], the MSE-minimising ∆2h ⋆ one has γ = ∆2 +τ 2 /N , which lies in [0, 1] and is monotone-decreasing in τ 2 . h
3.3
The DISCA estimator
Proposition 1 tells us what shape the shrinkage should have but not how to compute it: γ ⋆ depends on the unobservable ∆h . In particular, the proposition’s two parts work together to motivate the 4
method. Part (ii) gives the optimal γ ⋆ in closed form, but it is not directly computable: γ ⋆ depends on τ 2 and on ∆h , neither of which we observe. Part (i) gives us a way to estimate τ 2 from data we already have: the within-agent variance D2 has expected value equal to τ 2 . As ∆h has no analogous estimator from the agents’ outputs, we cannot compute γ ⋆ exactly. Instead, we construct an estimator δ ⋆ that is computable from the agents’ outputs alone, matches the shape of γ ⋆ , bounded in (0, 1), & monotone-decreasing in the variance estimate. We use N =4 agents per target country: three age cohorts (young, middle-aged, older) plus a country-wide aggregate, each instantiated as a native-language system prompt built from cohortand country-level aggregates of the WVS cultural profile of Greco et al. [2026] (full prompts in App. A12). The choice N =4 is the smallest panel that simultaneously covers the three demographic axes documented to carry most within-country variance [Inglehart and Welzel, 2005] and includes a country-wide aggregate that anchors the ensemble against cohort-specific sampling noise; the marginal precision gain from each additional persona decays as Θ(N −2 ) (Corollary 1, App. A2), placing N =4 on the flat side of the diminishing-returns curve. The empirical sweep over N ∈ {2, 3, 4, 5, 6} on a three-country panel (Table 22) confirms this: N =4 minimises macro MIS, N =2 leaves the consensus under-determined, and N ≥5 duplicates demographic coverage without adding new variance. To remove the order bias of forced-choice decoding, we evaluate every scenario under both (A, B) and (B, A) orderings and take the order-symmetrised gap (AB) (BA) − δi ) as the agent’s output. δi = 21 (δi 3.3.1
Loss-averse importance sampling
The plain consensus adjustment ∆ = δ̄ − δbase moves the base model toward the average direction because δ̄ is the average consensus. The problem is that an average direction can hide a cohort that strongly disagrees: three cohorts mildly in favour and one cohort strongly against will average out to "mildly in favour," and the consensus will move the model in a direction that one cohort actively rejects. We want a steering rule that avoids this. One that prefers a direction nobody strongly dislikes over a direction with a higher average but a strong dissenter. We frame this as a cooperative-bargaining problem among the cohorts [Nash, 1950, Kalai and Smorodinsky, 1975]. Each cohort is a player with its own utility over correction directions, each is loss-averse, which means a small misalignment with one’s preferred direction is felt more than a same-sized improvement, and we want to pick a direction that all players can collectively accept. This framing organises three design choices: where to look for candidates, how each player scores a candidate, and how all scores combine into a group verdict. Rather than committing to ∆, we draw K candidate perturbations ϵk ∼ N (0, σ 2 ) in a neighbourhood of the consensus, defining the candidate state δ̃k = δ̄ + ϵk . Each candidate is scored by how much closer to cohort i’s preferred direction it lands, in absolute ℓ1 distance, against two reference points: the cohort gap δi and the panel consensus δ̄, gi,k = |δbase − δi | − |δ̃k − δi |,
gcons,k = |δbase − δ̄| − |δ̃k − δ̄|,
(1)
so positive g means the candidate moved closer to the relevant target than the base prompt did, negative means it moved the model away. We need a scoring rule v(·) that is concave on the gain side so additional gain beyond what’s already aligned is worth less and steeper on the loss side so any cohort loss matters more than the same-size gain elsewhere. The canonical such rule in the decision-theory literature is the Kahneman–Tversky value function [Kahneman and Tversky, 1979, Tversky and Kahneman, 1992]: α z z ≥ 0, v(z) = (2) −κ (−z)α z < 0, with curvature α ∈ (0, 1] and loss-aversion κ ≥ 1. The asymmetric kernel v acts as an aggregator that penalises misalignment more than rewarding incremental alignment. Cooperative bargaining sits between two extremes: a purely utilitarian verdict averages the players’ utilities and ignores how the gains are distributed, while a purely collective verdict asks how the deal serves the group as a whole. We use a blend of the two: N
g 1 X gi,k cons,k v + λcoop v , Utotal (ϵk ) = (1 − λcoop ) N i=1 σ σ 5
(3)
where gcons,k measures the deal’s gain against the consensus direction δ̄ rather than against any single cohort. The first term is the utilitarian verdict over loss-averse players; the second is the collective verdict scored by the same loss-averse rule. The coefficient λcoop ∈ [0, 1] controls the bargaining position, with λcoop = 0 purely utilitarian and λcoop = 1 purely collective. We use λcoop = 0.7, on the collective side; a sensitivity sweep over λcoop ∈ [0, 1] is in App. A11. The division by σ inside v scores each gain in units of the perturbation scale, so the search radius σ and the value-function shape (α, κ) can be set independently. The natural choice, returning the highest-utility candidate, is unsafe here. The utility Utotal (ϵk ) is itself a noisy quantity. It depends on the random draw ϵk and on the agent gains gi,k , both of which carry stochastic error. Picking the argmax over K noisy scores selects the candidate that drew favourable noise on this sample, not the candidate with the highest true utility. We therefore aggregate all candidates with weights proportional to their utility: PK k=1 wk ϵk , wk ∝ exp Utotal (ϵk )/η , (4) δP T −IS = P K k=1 wk where η > 0 is an IS softmax temperature (η=0.5 default; Table 19). This form is not ad hoc: it is the path-integral solution to the stochastic optimal control problem of selecting a correction direction under a noisy utility function [Williams et al., 2018, Levine, 2018]. The high-utility candidate dominates the average, but near-winners contribute as well, and stochastic error in any one candidate’s utility is suppressed by the contributions of the others. We refer to the full procedure as loss-averse importance sampling, or PT-IS for short. 3.3.2
Dual-pass reliability gate
PT-IS gives us a single correction δPT-IS for each scenario. The natural next step is to apply it: add δPT-IS to the model’s base logits and read off the steered decision. The problem is that δPT-IS is a single number combining cohort opinions. When all cohorts agree, that number reflects a unanimously backed direction. When the cohorts disagree, the same number can come out small even though no cohort actually wants the small push it represents. Concretely, imagine two cohorts prefer direction +0.5 and two prefer −0.5. PT-IS with lossaversion will not pick either side; it returns a compromise, say δPT-IS = +0.05, close to zero because pushing in either direction would hurt half the cohorts. Applying +0.05 to the model’s logits nudges the model in a direction no cohort actually asked for. This is the failure mode Proposition 1 warned us about: γ ⋆ contracts toward zero when cohort disagreement is high, because the appropriate response to disagreement is to apply less correction, not a more carefully averaged correction. We need a way to identify contested scenarios at run time and contract δPT-IS accordingly. The obvious approach is to increase K. With infinitely many candidates, δPT-IS converges to a deterministic limit and run-to-run variation vanishes. But this does not solve the problem: the +0.05 compromise becomes a more precise compromise. More samples sharpen the average; they do not change whether the average deserves to be applied at full strength. We run PT-IS twice on the same scenario, on disjoint sub-budgets of size K/2, producing indepen(1) (2) dent corrections δPT-IS and δPT-IS . When the cohorts agree, both runs find the same compromise and the two corrections are close. When the cohorts disagree, the runs are pulled by competing cohort camps and land on different compromises. The squared inter-run gap (1)
(2)
Vr = (δPT-IS − δPT-IS )2 is the diagnostic: zero when the runs agree, growing as they diverge. Vr is the half-sample variance estimator: a single sample is split in two, and variance is estimated from the squared inter-half gap. The technique originates in Mahalanobis [Mahalanobis, 1946] and was generalised as balanced repeated replication [McCarthy, 1969, Wolter, 2007]; it is canonical in settings that need a qualitative variance signal on a constrained compute budget. Two runs also keep total compute matched to a single full PT-IS run. We multiplicatively shrink the correction by r = exp −Vr /s ∈ (0, 1],
(5)
with bandwidth s > 0 controlling the decay rate. The exponential is the simplest function satisfying the three shape properties Proposition 1 requires: bounded in (0, 1], monotone-decreasing, and 6
smooth. The final DISCA correction is the gated average of the two runs: (1)
δ⋆ = r ·
(2)
δPT-IS + δPT-IS . 2
(6)
δ ⋆ matches the qualitative shape of γ ⋆ · ∆: bounded by the average run magnitude, monotonedecreasing in inter-run disagreement (mirroring γ ⋆ in τ 2 ), and smooth in all inputs. The empirical contribution of the gate is bundled inside the full-method ablation in Table 4. 3.4
Extension to Open-Ended Ethical Scenarios
To extend DISCA to open-ended scenarios, the model generates free-form responses under both the original prompt and four persona-specific prompts. A separate large language model (LLM) judge then parses these outputs to return a {choice, confidence} pair. The continuous decision signal is represented by a pseudo-logit gap (the A/B logit distance) extracted by that judge. The dual-pass Prospect-Theory importance sampling (PT–IS) mechanism is applied similarly to the binary case.
4
Experimental Results
We test DISCA on MultiTP [Jin et al., 2025], the multilingual extension of the Moral Machine trolley problem [Awad et al., 2018]. Each scenario shows the model two groups of people that an autonomous vehicle could spare, and the model has to choose one of them by emitting a single decision token, A or B. Every scenario isolates one of six moral attributes (species, gender, age, fitness, social value, and the number of lives saved) while matching the two groups on everything else. An age scenario, for example, pits a young group against an older one and holds the rest of the description constant; the rate at which the model spares the younger group, averaged across many such scenarios, is then read off as its preference on age. Following the Moral Machine convention we always assign decision token B to the side the AMCE measures toward; this is a measurement convention from conjoint analysis that fixes the sign of the preference, not a normative claim about which side ought to be saved. The frozen model emits logits z(x) = [za , zb ] at the decision position, the decision gap is δ(x) = zb − za , and the sparing probability under decoding temperature Tdec is pspare (x) = σ(δ(x)/Tdec ). Averaging the sparing probability across the scenarios for one attribute and one country gives that country’s Average Marginal Component Effect (AMCE): X 1 (c) m̂d = (c) pspare (x). (7) |Sd | (c) x∈Sd
Stacking the six attributes gives a six-dimensional vector m̂(c) that we compare to the human AMCE vector h(c) aggregated from real Moral Machine judgments [Awad et al., 2018, Jin et al., 2025]. Our headline metric is the misalignment score MIS(c) = ∥m̂(c) − h(c) ∥2 , which is just the straight-line distance between the model’s preference vector and the country’s human one; lower is better. We report it macro-averaged over a 20-country panel that spans four continents. The country list, the preprocessing pipeline (deduplication, quality filter, per-attribute caps, oversampling, and deterministic shuffling), the per-country slice sizes, and a cap-disabled sensitivity rerun all live in App. A14. Every scenario is presented in the country’s native language; only the decision tokens A and B stay in English so that token-ID extraction is consistent across model families. We also report Pearson r for shape agreement and Jensen–Shannon divergence (JSD) as secondary diagnostics. We tested 28 model–method combinations across 12 architectures from 270M to 70B parameters, with the full landscape in App. A7, and report seven open-weight backbones with the most robust gains: Llama-3.3-70B, Magistral-24B, Phi-4 (14B), Qwen3-VL-8B, Qwen2.5-7B, Phi-3.5-mini (3.8B), and Gemma-4-E2B (2B). The few models that degrade share a single pattern: their vanilla MIS is already low enough that there is no headroom left for any correction (we return to this in §4.2). The released hyperparameter configuration was validated on a small three-backbone, fivecountry prototyping panel before being frozen for the headline sweep; how we elicit the A and B tokens consistently across model families is described in App. A13. We compare DISCA against two tiers of baselines on the same 20-country Phi-4 grid (Table 1; full implementations in App. A10). The first tier is training-free and never sees the human AMCE: vanilla decoding, a WVS Profile Prompt that summarises the country’s WVS values, a PRISM-style cultural framing prompt [Kirk et al., 2024], a fixed logit offset, activation steering [Arditi et al., 2025, Zou et al., 2023], and MC-Dropout [Gal and Ghahramani, 2016, Kwon et al., 2026]. We exclude 7
Table 1: Inference-time baseline comparison on the Phi-4 20-country grid (macro MIS, lower is better). Oracle baselines (below dashed line) use human AMCE during calibration; even so they fail to match DISCA. AMCE?
MIS ↓
DISCA (ours) PRISM-Style Prompt [Kirk et al., 2024] MC-Dropout [Gal and Ghahramani, 2016] Activation steering [Arditi et al., 2025] Fixed logit offset WVS Profile Prompt Vanilla decoding
No No No No No No No
0.346 0.384 0.403 0.430 0.439 0.453 0.454
Margin scaling [Guo et al., 2017] Temp. scaling [Guo et al., 2017]
Yes Yes
0.506 0.513
Method
Table 2: Macro results on the 20-country MultiTP slice (equal weight per country). MIS is mean ℓ2 AMCE distance (± std over three seeds {42, 101, 2026} chosen under a fixed compute budget; per-cell std median 0.008, full multiseed analysis in §A9.1); “Win” = number of countries with lower DISCA MIS than vanilla. Model Llama-3.3-70B Magistral-24B Phi-4 (14B) Qwen3-VL-8B Qwen2.5-7B Phi-3.5-mini Gemma-4-E2B (2B)
MIS ↓
Gain (%)
Win/20
r
.668±.006 .350±.005 .346±.004 .466±.006 .362±.005 .571±.006 .479
+21.3±0.8 +13.0±0.7 +23.6±0.6 +18.8±0.9 +20.0±0.8 +10.3±1.0 +3.4
20 16 18 20 17 17 14
−0.01 +0.62 +0.56 +0.22 +0.45 −0.35 +0.46
reward-guided decoding [Khanov et al., 2024, Mudgal et al., 2024] because it assumes per-country reward models, which simply do not exist for the long tail of countries we care about. The second tier is the opposite: oracle baselines that fit one per-country scalar (temperature or additive margin) directly on the country’s full human AMCE [Guo et al., 2017], serving as upper bounds on what any 1-d-per-country oracle correction can do.
4.1
Main results
A single decoding-time intervention, with no weight changes and no per-country labels, materially closes the gap between large language models and human moral preferences. Across seven backbones from five model families, DISCA reduces macro MIS by roughly ten to twenty-four percent relative to vanilla decoding (Table 2), and two of the seven backbones (Llama-3.3-70B and Qwen3VL-8B) improve on every one of the twenty countries we test. A 2D PCA of the AMCE vectors tells the same story visually (Figure 3, App. A6): every country point migrates from the vanilla cloud toward the human cloud, none in the wrong direction. No simpler intervention comes close. On the same Phi-4 grid (Table 1) DISCA outperforms every training-free baseline by a clear margin, and even the two oracle baselines that get to peek at each country’s full human AMCE during calibration end up worse than DISCA, because a single per-country scalar (whether a temperature or an additive margin) cannot capture all six AMCE dimensions simultaneously. The per-country breakdown (Table 7, App. A4) confirms that the macro gains are not concentrated in a handful of lucky cells. Geographic patterns, large single-country swings, and the harder regions (such as those with sparser WVS coverage) are all detailed there. To rule out artefacts, we also ran nine independent robustness checks and three random seeds; every conclusion sits comfortably within the bootstrap noise floor (Appendices A11 and A9.1). Calibration competes with scale The single most striking finding is a crossing in the curve of MIS against model size (App. A6): a 14B model with DISCA (Phi-4) reaches lower absolute misalignment than a 70B model without it (Llama-3.3-70B). Per-dimension error analysis (Table 10, App. A5) explains why. Different models hit a wall on different attributes. Weaker models are dominated by Utilitarianism errors that are simply too large for any persona-based correction to close. Better-calibrated models have already resolved Utilitarianism and Species, so their bottleneck shifts to Social Value and Age, two attributes that DISCA’s importance-sampling stage actually moves the needle on. Phi-4 turns out to be the most balanced of the seven across all dimensions, and this, rather than raw parameter count, brings its top performance despite being five times smaller than the baseline. Generalisation beyond binary dilemmas The disagreement signal extends to open-ended ethical generation (§3.4). All four backbones improve on average, with gains up to ∼7% (Table 3); only six of eighty cells regress (Table 8), all with |∆| ≤ 3.1%—the gate bounds but cannot reverse them. A logit-noise stress test (App. A4.2) shows the gated pipeline retains a 10–12% MIS advantage over an ungated variant (the same PT–IS stack without agreement-based shrinkage) across σ ∈ [0, 2], with the gap widening as noise grows. 8
Table 3: Open-ended track summary (20 countries, 310 scenarios each). VAN = vanilla decoding (no DISCA). DISCA = full pipeline. MIS = misalignment score (§4); lower MIS is better. ∆% reports mean relative MIS reduction across four backbones, so higher ∆% is better (positive = improvement); per-country breakdown in App. A4.1.
4.2
Model
Vanilla
DISCA
∆%
Llama-3.3-70B Phi-4 (14B) Qwen2.5-7B Phi-3.5-mini
0.471 0.324 0.321 0.521
0.439 0.302 0.306 0.510
+6.85% +6.67% +4.55% +2.13%
Ablation and Discussion
DISCA works across families, geographies, and formats. The remaining questions are why, where it stops, and what it cannot do. Ablating four design choices on three backbones (Table 4)1 gives a consistent ranking. Persona diversity is the irreplaceable core: removing the WVS personas hurts 16–18 of 20 countries per backbone—the largest single-component loss—confirming that withincountry disagreement carries most of the alignment value. PT–IS adds a stable second-order gain; the reliability gate matters more than raw correction strength; positional debiasing has the smallest macro effect but remains necessary as a conditioning step. Table 4: Cross-backbone ablation on all 20 paper countries. Lower MIS is better (↓ in column headers); ∆ vs Full reports the change in MIS relative to Full DISCA, so positive values indicate worse alignment. “n hurt” = number of countries with strictly higher MIS than Full DISCA. Qwen2.5-7B (BF16) Configuration
MIS ↓
∆ vs Full
n hurt
Phi-3.5-mini-Instruct MIS ↓
∆ vs Full
n hurt
Magistral-Sml (24B) MIS ↓
∆ vs Full
n hurt
Full DISCA
0.362
–
–
0.571
–
–
0.334
–
–
Without persona Always-on PT–IS No-IS (consensus) No debiasing
0.417 0.379 0.371 0.363
+.055 +.017 +.009 +.001
16 14 12 12
0.618 0.586 0.579 0.573
+.047 +.015 +.008 +.002
17 14 13 11
0.395 0.355 0.345 0.338
+.061 +.021 +.011 +.004
18 15 13 10
DISCA also knows when not to act. The dual-pass gate passes most corrections through unattenuated and shrinks aggressively when the two passes disagree (audit in App. A9). Replacing the bounded, loss-averse aggregation with simple averaging barely moves the mean but nearly quadruples the harmed cells and triples worst-case degradation (Table 15, App. A9.2): the bounded form is a safety mechanism, not an average-case booster. Failures are architectural, not cultural. The 28-model sweep (App. A7) shows success tracks logit conditioning, not parameter count: Phi-4 leads at 14B while several 30B+ models degrade [Chand et al., 2026]. Per-scenario decision margin, entropy, and absolute logit gap (App. A9) flag likely failures from the vanilla pass alone, so a deployment-time check can skip poorly-conditioned models. The price is roughly 4× vanilla latency (App. A15). The signal has a clear scope boundary: on BLEnD factual cultural QA [Myung et al., 2024] (App. A8) it does not help. MultiTP value alignment lives in a scalar logit gap that importance sampling navigates well; factual QA picks one token from a ∼32k-entry vocabulary, where perturbations tuned for a binary gap become noise on a wide softmax. Drawing this line between value steering and fact retrieval is itself a contribution. Three caveats bound generalisation. We optimise against crowdsourced AMCE vectors [Awad et al., 2018, Jin et al., 2025] without a separate study of perceived appropriateness [Atari et al., 2023, Khan et al., 2025], so we measure alignment to a survey statistic, not legitimacy. The Prospect-Theory function is an aggregation kernel, not a cognitive model. DISCA needs decision-token logits, which black-box text-only APIs do not always expose (App. A17). An inference-time controller can also encode harmful majorities if deployed naively [Zewail et al., 2026]; a per-persona utility floor caps how far any persona’s post-correction utility may drop below vanilla, and sweeping the floor leaves 1 vLLM backend, one K-sample IS batch per scenario. Full DISCA keeps the dual-pass gate; the gate is disabled only for rows downstream of it (Always-on PT–IS, No-IS (consensus)), so those rows quantify what the gate hides without altering the headline.
9
macro MIS flat within the bootstrap noise (Table 25, App. A11), so the safeguard is essentially free.
5
Conclusion
In this paper we have shown that LLMs need not be retrained to respect moral diversity. To do so, we have introduced disagreement-driven steering, a paradigm that treats within-group variance of grounded agents as the primary alignment signal rather than their consensus. DISCA instantiates this paradigm for cultural alignment, reducing misalignment by 10–24% on the six headline backbones (≥3.8B), and 2–7% on open-ended scenarios, across twenty countries without changing a single weight. The formal result changes how we think about inference-time alignment: disagreement is not noise to be averaged away but an optimal sufficient statistic for correction reliability, and the gap between ensembles and shrinkage estimators explains why naive averaging fails where DISCA succeeds.
References M. S. Z. b. Ahmad and K. Takemoto. Large-scale moral machine experiment on large language models. PLOS ONE, 20(5):e0322776, 2025. doi: 10.1371/journal.pone.0322776. URL https: //doi.org/10.1371/journal.pone.0322776. A. Arditi, O. Obeso, A. Syed, D. Paleka, N. Panickssery, W. Gurnee, and N. Nanda. Refusal in language models is mediated by a single direction. In Advances in Neural Information Processing Systems, volume 37, pages 136037–136083, 2025. doi: 10.52202/079017-4322. URL https://proceedings.neurips.cc/paper_files/paper/ 2024/hash/f545448535dfde4f9786555403ab7c49-Abstract-Conference.html. M. Atari, M. J. Xue, P. S. Park, D. E. Blasi, and J. Henrich. Which humans? PsyArXiv preprint, 2023. doi: 10.31234/osf.io/5b26t. https://osf.io/preprints/psyarxiv/5b26t. E. Awad, S. Dsouza, R. Kim, J. Schulz, J. Henrich, A. Shariff, J.-F. Bonnefon, and I. Rahwan. The moral machine experiment. Nature, 563(7729):59–64, 2018. S. Chand, F. Baca, and E. Ferrara. No free lunch in language model bias mitigation? Targeted bias reduction can exacerbate unmitigated LLM biases. AI, 7(1):24, 2026. doi: 10.3390/ai7010024. R. Chen, W. Chai, Z. Yang, X. Zhang, Z. Wang, T. Quek, J. T. Zhou, S. Poria, and Z. Liu. DiffPO: Diffusion-styled preference optimization for inference time alignment of large language models. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 18910–18925, Vienna, Austria, July 2025. Association for Computational Linguistics. doi: 10.18653/v1/2025.acl-long.926. URL https: //aclanthology.org/2025.acl-long.926/. arXiv:2503.04240. Y. Du, S. Li, A. Torralba, J. B. Tenenbaum, and I. Mordatch. Improving factuality and reasoning in language models through multiagent debate. In Proceedings of the 41st International Conference on Machine Learning, volume 235 of Proceedings of Machine Learning Research, pages 11733–11763. PMLR, 2024. URL https://proceedings.mlr.press/v235/du24e.html. arXiv:2305.14325. K. Ethayarajh, W. Xu, N. Muennighoff, D. Jurafsky, and D. Kiela. KTO: Model alignment as prospect theoretic optimization. In International Conference on Machine Learning, 2024. arXiv:2402.01306. Y. Gal and Z. Ghahramani. Dropout as a Bayesian approximation: Representing model uncertainty in deep learning. In Proceedings of the 33rd International Conference on Machine Learning, volume 48 of Proceedings of Machine Learning Research, pages 1050–1059. PMLR, 2016. URL https://proceedings.mlr.press/v48/gal16.html. C. M. Greco, L. La Cava, and A. Tagarelli. Culturally grounded personas in large language models: Characterization and alignment with socio-psychological value frameworks. arXiv preprint arXiv:2601.22396, 2026. 10
C. Guo, G. Pleiss, Y. Sun, and K. Q. Weinberger. On calibration of modern neural networks. In Proceedings of the 34th International Conference on Machine Learning, volume 70 of Proceedings of Machine Learning Research, pages 1321–1330, 2017. URL https://proceedings.mlr. press/v70/guo17a.html. arXiv:1706.04599. C. Haerpfer, R. Inglehart, A. Moreno, C. Welzel, K. Kizilova, J. Diez-Medrano, M. Lagos, P. Norris, E. Ponarin, and B. Puranen. World Values Survey: Round seven – country-pooled datafile. Madrid, Spain & Vienna, Austria: JD Systems Institute & WVSA Secretariat, 2020. https://doi.org/10.14281/18241.20. J. Henrich, S. J. Heine, and A. Norenzayan. The weirdest people in the world? Behavioral and Brain Sciences, 33(2-3):61–83, 2010. doi: 10.1017/S0140525X0999152X. URL https://www.cambridge.org/core/journals/behavioral-and-brain-sciences/ article/abs/weirdest-people-inthe-world/BF84F7517D56AFF7B7EB58411A554C17. R. Inglehart and C. Welzel. Modernization, Cultural Change, and Democracy: The Human Development Sequence. Cambridge University Press, 2005. URL https://social.hse.ru/data/ 2012/11/03/1249193128/inglehart_welzel.pdf. Z. Jin, M. Kleiman-Weiner, G. Piatti, S. Levine, J. Liu, F. G. Adauto, F. Ortu, A. Strausz, M. Sachan, R. Mihalcea, Y. Choi, and B. Schölkopf. Language Model Alignment in Multilingual Trolley Problems. In International Conference on Learning Representations, 2025. URL https:// openreview.net/forum?id=VEqPDZIDAh. arXiv:2407.02273. D. Kahneman and A. Tversky. Prospect theory: An analysis of decision under risk. Econometrica, 47(2):263–291, 1979. E. Kalai and M. Smorodinsky. Other solutions to Nash’s bargaining problem. Econometrica, 43(3): 513–518, 1975. doi: 10.2307/1914280. A. Khan, S. Casper, and D. Hadfield-Menell. Randomness, not representation: The unreliability of evaluating cultural alignment in LLMs. In Proceedings of the 2025 ACM Conference on Fairness, Accountability, and Transparency, pages 2151–2165. Association for Computing Machinery, 2025. doi: 10.1145/3715275.3732147. URL https://dl.acm.org/doi/10.1145/ 3715275.3732147. arXiv:2503.08688. M. Khanov, J. Burapacheep, and Y. Li. ARGS: Alignment as reward-guided search. In International Conference on Learning Representations, 2024. URL https://openreview.net/forum?id= shgx0eqdw6. arXiv:2402.01694. J. Kim, J. Kwon, L. F. Vecchietti, A. Oh, and M. Cha. Exploring persona-dependent llm alignment for the moral machine experiment. arXiv preprint arXiv:2504.10886, 2025. doi: 10.48550/arXiv. 2504.10886. URL https://arxiv.org/abs/2504.10886. H. R. Kirk, A. Whitefield, P. Röttger, A. Bean, K. Margatina, J. Ciro, R. Mosquera, M. Bartolo, A. Williams, H. He, B. Vidgen, and S. A. Hale. The PRISM alignment dataset: What participatory, representative and individualised human feedback reveals about the subjective and multicultural alignment of large language models. arXiv preprint arXiv:2404.16019, 2024. doi: 10.48550/arXiv.2404.16019. URL https://arxiv.org/abs/2404.16019. The Thirty-eighth Conference on Neural Information Processing Systems, Datasets and Benchmarks Track (2024). J. Kwon, L. F. Vecchietti, S. Park, and M. Cha. Dropouts in confidence: Moral uncertainty in human-LLM alignment. In Proceedings of the AAAI Conference on Artificial Intelligence, 2026. arXiv:2511.13290. S. Levine. Reinforcement learning and control as probabilistic inference: Tutorial and review. arXiv preprint arXiv:1805.00909, 2018. doi: 10.48550/arXiv.1805.00909. URL https://arxiv. org/abs/1805.00909. T. Liang, Z. He, W. Jiao, X. Wang, Y. Wang, R. Wang, Y. Yang, S. Shi, and Z. Tu. Encouraging divergent thinking in large language models through multi-agent debate. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 17889–17904. Association for Computational Linguistics, 2024. doi: 10.18653/v1/2024.emnlp-main.992. URL https://aclanthology.org/2024.emnlp-main.992/. 11
P. C. Mahalanobis. Recent experiments in statistical sampling in the Indian Statistical Institute. Journal of the Royal Statistical Society, 109(4):325–378, 1946. P. J. McCarthy. Pseudo-replication: Half samples. Review of the International Statistical Institute, 37(3):239–264, 1969. S. Mudgal, J. Lee, H. Ganapathy, Y. Li, T. Wang, Y. Huang, Z. Chen, H.-T. Cheng, M. Collins, T. Strohman, J. Chen, A. Beutel, and A. Beirami. Controlled decoding from language models. In Proceedings of the 41st International Conference on Machine Learning, volume 235 of Proceedings of Machine Learning Research, pages 36486–36503. PMLR, 2024. URL https://proceedings.mlr.press/v235/mudgal24a.html. arXiv:2310.17022; OpenReview: nrEqH502eC. J. Myung, N. Lee, Y. Zhou, J. Jin, R. A. Putri, D. Antypas, H. Borkakoty, E. Kim, C. PerezAlmendros, A. A. Ayele, et al. BLEnD: A benchmark for LLMs on everyday knowledge in diverse cultures and languages. In Advances in Neural Information Processing Systems (NeurIPS), Datasets and Benchmarks Track, 2024. doi: 10.48550/arXiv.2406.09948. URL https://openreview.net/forum?id=nrEqH502eC. arXiv:2406.09948. J. F. Nash. The bargaining problem. Econometrica, 18(2):155–162, 1950. doi: 10.2307/1907266. M. Rudelson and R. Vershynin. Hanson–Wright inequality and sub-Gaussian concentration. Electronic Communications in Probability, 18(82):1–9, 2013. doi: 10.1214/ECP.v18-2865. M. J. Ryan, W. Held, and D. Yang. Unintended impacts of LLM alignment on global representation. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 16121–16140, Bangkok, Thailand, Aug. 2024. Association for Computational Linguistics. doi: 10.18653/v1/2024.acl-long.853. URL https: //aclanthology.org/2024.acl-long.853/. arXiv:2402.15018. S. Santurkar, E. Durmus, F. Ladhak, C. Lee, P. Liang, and T. Hashimoto. Whose opinions do language models reflect? arXiv preprint arXiv:2303.17548, 2023. T. Sorensen, J. Moore, J. Fisher, M. Gordon, N. Mireshghallah, C. M. Rytting, A. Ye, L. Jiang, X. Lu, N. Dziri, et al. A roadmap to pluralistic alignment. arXiv preprint arXiv:2402.05070, 2024. doi: 10.48550/arXiv.2402.05070. URL https://arxiv.org/abs/2402.05070. K. Takemoto. The moral machine experiment on large language models. Royal Society Open Science, 11(2):231393, 2024. doi: 10.1098/rsos.231393. URL https:// royalsocietypublishing.org/rsos/article/11/2/231393/92489. A. M. Turner, L. Thiergart, G. Leech, D. Udell, J. J. Vazquez, U. Mini, and M. MacDiarmid. Activation addition: Steering language models without optimization. arXiv preprint arXiv:2308.10248, 2024. doi: 10.48550/arXiv.2308.10248. URL https://arxiv.org/abs/2308.10248. A. Tversky and D. Kahneman. Advances in prospect theory: Cumulative representation of uncertainty. Journal of Risk and Uncertainty, 5(4):297–323, 1992. doi: 10.1007/BF00122574. URL https://link.springer.com/article/10.1007/BF00122574. X. Wang, J. Wei, D. Schuurmans, Q. V. Le, E. H. Chi, S. Narang, A. Chowdhery, and D. Zhou. Selfconsistency improves chain of thought reasoning in language models. In International Conference on Learning Representations (ICLR), 2023. doi: 10.48550/arXiv.2203.11171. URL https: //openreview.net/forum?id=1PL1NIMMrw. arXiv:2203.11171. G. Williams, P. Drews, B. Goldfain, J. M. Rehg, and E. A. Theodorou. Information-theoretic model predictive control: Theory and applications to autonomous driving. IEEE Transactions on Robotics, 34(6):1603–1622, 2018. doi: 10.1109/TRO.2018.2865891. URL https: //ieeexplore.ieee.org/abstract/document/8558663. K. M. Wolter. Introduction to Variance Estimation. Springer, 2nd edition, 2007. J. Yao, X. Yi, J. Wang, Z. Dou, and X. Xie. CAReDiO: Cultural alignment of LLM via representativeness and distinctiveness guided data optimization. arXiv preprint arXiv:2504.08820, 2025. doi: 10.48550/arXiv.2504.08820. URL https://arxiv.org/abs/2504.08820. 12
A. Zewail, A. Figueroa, J. Graham, and M. Atari. Moral stereotyping in large language models. Proceedings of the National Academy of Sciences, 123(10):e2519941123, 2026. doi: 10.1073/ pnas.2519941123. URL https://www.pnas.org/doi/10.1073/pnas.2519941123. B. Zhang, X. Zhao, J. Li, H. Chen, and Z. Chen. Mind the gap in cultural alignment: Task-aware culture management for large language models. arXiv preprint arXiv:2602.22475, 2026. A. Zou, L. Phan, S. Chen, J. Campbell, P. Guo, et al. Representation engineering: A top-down approach to AI transparency. arXiv preprint arXiv:2310.01405, 2023.
13
Appendix Contents A1
MultiTP Dimensions and DISCA Inference Pseudocode
16
A2
Proof of Proposition 1
16
A3
Additional Theoretical Properties of DISCA
19
A4
A5
A3.1
Bounded-correction safety guarantee . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
A3.2
Finite-sample concentration of the within-panel variance estimator . . . . . . . . . . . 21
A3.3
Hölder stability with respect to backbone perturbations . . . . . . . . . . . . . . . . . . . . 23
Full Per-Country Results
25
A4.1
Full Open-Ended Per-Country Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
A4.2
Logit-Noise Robustness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
Per-Dimension Error Analysis A5.1
28
Diagnosing negative Pearson r . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
A6
Scaling and Geographic Visualizations
30
A7
Broader Model Landscape
31
A8
Factual Cultural QA Evaluation (BLEnD)
31
A9
Post-Hoc Diagnostics
32
A9.1
Multi-seed stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
A9.2
Step 3 is a tail-safety mechanism . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
A9.3
Scenario-level conditioning of the DISCA correction . . . . . . . . . . . . . . . . . . . . . . 33
A10 Baseline Implementation Details
33
A10.1 Other training-free baselines (Table 1) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34 A10.2 Activation Steering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34 A10.3 ARGS and controlled decoding (relation to our work) . . . . . . . . . . . . . . . . . . . . . 34 A10.4 Additional inference-time baselines (twenty-country Phi-4 grid) . . . . . . . . . . . . . 34 A11 Hyperparameters and Sensitivity Sweeps
36
A11.1 Robustness Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36 A11.2 Default Hyperparameters and Validation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36 A11.3 Temperature Sensitivity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37 A11.4 Extended Hyperparameter Sensitivity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37 A12 Persona Construction Details
39
A12.1 Persona Prompts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39 A12.2 WVS Data Processing Pipeline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42 A12.3 WVS-to-Trolley Dimension Linkage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42 A12.4 Macro-level WVS dimension importance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43 A12.5 Sensitivity to the Fourth Persona . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44 A13 AMCE Estimation Details
44 14
A14 Dataset Preprocessing and Sensitivity
45
A15 Trigger Mechanism and Latency
46
A16 Relationship to Persona-Dependent LLM Alignment
47
A17 Extended Limitations
48
15
A1
MultiTP Dimensions and DISCA Inference Pseudocode
The pipeline schematic is Figure 1 in §3. Table 5 lists the six MultiTP moral dimensions and the per-attribute logit temperatures Tc used by DISCA (validated in App. A11). Algorithm 1 states the exact inference routine. Algorithm 1 DISCA Inference: Scenario x in Country c Require: Frozen LLM fθ ; personas {si }N i=1 ; hyperparameters Khalf , σ, λ, η, Tdec , s Ensure: pspare ∈ [0, 1] 1: for order ∈ {orig, swap} do 2: Batched forward: fθ on base + N personas (ord) 3: Extract logit pairs; compute δi 4: end for (orig) (swap) 5: δi ← (δi − δi )/2 P 1 6: Compute consensus δ̄ ← N i δi 7: for m ∈ {1, 2} do 8: for k = 1 to Khalf do 9: Sample ϵk ∼ N (0, σ 2 ); compute Utotal (ϵk ) via Eqs. 1, 3 10: end forP (m) ⋆ 11: δm ← k wk ϵk if ESSm > ρeff else 0 12: end for 13: r ← exp(−(δ1⋆ − δ2⋆ )2 /s) 14: δ ⋆ ← r (δ1⋆ + δ2⋆ )/2 15: δfinal ← αess δ̄ + (1 − αess ) δbase + δ ⋆ (with αess = min(1, ESS/ρeff )) 16: return pspare = σ(δfinal /Tdec )
Table 5: The six moral dimensions in MultiTP. Tc is the per-attribute logit temperature used by our method; values reflect empirical logit magnitudes and are validated in App. A11.
A2
Dimension
Preferred
Contra
Tc
Species Gender Age Fitness Social Value Utilitarianism
Human Female Young Fit High (exec.) More lives
Animal Male Elderly Unfit Low (homeless) Fewer lives
4.0 3.5 1.5 1.5 1.5 1.5
Proof of Proposition 1
Recall the agent model from §3.2: for a single scenario, δi = δh + ηi ,
i = 1, . . . , N,
where δh is a fixed (but unknown) population-preference logit gap, and the noise terms ηi are i.i.d. with E[ηi ] = 0 and Var(ηi ) = τ 2 . We also fix δbase , the gap from the base prompt (deterministic given the scenario), and define N
δ̄ =
1 X δi , N i=1
D2 =
N 1 X (δi − δ̄)2 , N − 1 i=1
∆ = δ̄ − δbase ,
∆h = δh − δbase .
Throughout, expectations and variances are taken over the noise η1 , . . . , ηN . Part (i): E[D2 ] = τ 2 and Var(∆) = τ 2 /N We first rewrite D2 in terms of the noise variables. For each i, X δi − δ̄ = (δh + ηi ) − δh + N1 ηj = ηi − η̄, j
16
where η̄ = N1 differences, so
P
j ηj .
The population term δh contributes equally to every δi and cancels in the D2 =
Expanding the square and using N X i=1
(ηi − η̄)2 =
N X
N 1 X (ηi − η̄)2 . N − 1 i=1
(8)
P
ηi2 − 2η̄
i=1
i ηi = N η̄, N N X X
ηi +
i=1
η̄ 2 =
i=1
N X
ηi2 − 2N η̄ 2 + N η̄ 2 =
i=1
N X
ηi2 − N η̄ 2 .
i=1
Taking expectations on both sides, " # X X 2 E (ηi − η̄) = E[ηi2 ] − N E[η̄ 2 ]. i
i
The two expectations on the right from the noise assumptions. For each i, E[ηi2 ] = P are immediate 2 2 2 2 Var(η ) + (E[η ]) = τ , so E[η ] = N τ . For η̄, linearity of expectation gives E[η̄] = i i i Pi 1 i E[ηi ] = 0, and independence of the ηi gives N 1 X 1 τ2 Var(η̄) = 2 Var(ηi ) = 2 · N τ 2 = , N i N N so E[η̄ 2 ] = Var(η̄) + (E[η̄])2 = τ 2 /N . Substituting, " # X τ2 E (ηi − η̄)2 = N τ 2 − N · = (N − 1)τ 2 , N i and dividing by N − 1 gives E[D2 ] = τ 2 .
(9)
The variance of the consensus adjustment follows directly. Since δbase is deterministic, τ2 Var(∆) = Var(δ̄ − δbase ) = Var(δh + η̄) = Var(η̄) = . N The two quantities just established play distinct roles: D2 has expected value τ 2 , the variance of one agent’s noise (this is what we estimate from data), whereas Var(∆) = τ 2 /N is the variance of the mean of N agents’ noise (this is what enters γ ⋆ in Part (ii)). Part (ii): MSE-minimising shrinkage We seek the value of γ ∈ [0, 1] that minimises h 2 i MSE(γ) = E γ∆ − ∆h . The strategy is to rewrite the error γ∆−∆h as a sum of a mean-zero random part and a deterministic constant, so that the cross term in the squared expansion vanishes. From Part (i), ∆ = δh + η̄ − δbase and E[η̄] = 0, so E[∆] = δh + 0 − δbase = ∆h . Adding and subtracting γ∆h inside the error, γ∆ − ∆h = (γ∆ − γ∆h ) + (γ∆h − ∆h ) = γ(∆ − ∆h ) + (γ − 1)∆h . Squaring and taking expectations, MSE(γ) = γ 2 E[(∆ − ∆h )2 ] + 2γ(γ − 1)∆h E[∆ − ∆h ] +(γ − 1)2 ∆2h | {z } =0
τ2 = γ2 · + (γ − 1)2 ∆2h , N where E[(∆ − ∆h )2 ] = Var(∆) = τ 2 /N from Part (i). The two surviving terms are the standard bias–variance pieces: 17
• γ 2 · τ 2 /N is the variance of γ∆ - it grows with how confidently we apply the consensus, scaled by the consensus’s own sampling variance. • (γ − 1)2 ∆2h is the squared bias from shrinkage - zero at γ = 1 (full adjustment, no bias), quadratic as γ → 0. Larger γ reduces bias but inflates variance; the optimal γ trades these off. To find γ ⋆ , differentiate MSE(γ): τ2 d MSE(γ) = 2γ · + 2(γ − 1)∆2h dγ N 2 τ 2 2 =2 γ + ∆h − ∆h . N Setting this to zero gives γ⋆ =
∆2h . ∆2h + τ 2 /N
(10)
The second derivative is 2(∆2h + τ 2 /N ) > 0 whenever τ 2 > 0, so γ ⋆ is the unique minimum. It remains to check that γ ⋆ behaves as the proposition claims. The numerator is non-negative and the denominator is at least as large, so γ ⋆ ∈ [0, 1] automatically. Writing γ ⋆ as a function of u := τ 2 /N , γ ⋆ (u) =
∆2h , ∆2h + u
dγ ⋆ ∆2 = − 2 h 2 ≤ 0, du (∆h + u)
so γ ⋆ is monotone-decreasing in the consensus variance, recovering the two boundary regimes: • τ 2 → 0 (agents agree): γ ⋆ → 1, apply the consensus in full. • τ 2 → ∞ or ∆h → 0 (signal-to-noise collapses): γ ⋆ → 0, do not adjust at all.
Corollary 1: Marginal Value of Additional Personas Corollary 1 (Marginal value of additional personas). Under the agent model of Proposition 1, the marginal reduction in consensus variance from adding the (N +1)-th persona is ∆Var(N → N +1) =
τ2 τ2 τ2 − = = Θ(N −2 ). N N +1 N (N + 1)
(11)
Equivalently, the optimal shrinkage coefficient γ ⋆ (N ) = ∆2h /(∆2h + τ 2 /N ) from Proposition 1(ii) satisfies γ ⋆ (N +1) − γ ⋆ (N ) = Θ(N −2 ) for N ≫ τ 2 /∆2h . Proof. Direct calculation from Var(∆) = τ 2 /N (Proposition 1(i)). The γ ⋆ statement follows from γ ⋆ (N +1) − γ ⋆ (N ) =
∆2h τ 2 , N (N + 1) (∆2h + τ 2 /N ) (∆2h + τ 2 /(N +1))
which decays as Θ(N −2 ) in the regime τ 2 /N ≪ ∆2h . Expressing the marginal precision gain as a fraction of τ 2 exposes the diminishing-returns structure: Table 6: Marginal reduction of Var(∆) from adding one persona, expressed as a fraction of the noise variance τ 2 . The drop from N =4 to N =5 already sits below 5%, while compute scales linearly in N. Transition
∆Var/τ 2
Fraction
N =2 → 3 N =3 → 4 N =4 → 5 N =5 → 6
1/6 1/12 1/20 1/30
16.7% 8.3% 5.0% 3.3%
18
Combined with the demographic-coverage requirement of N ≥ 3 (three age cohorts) plus a countrywide aggregate as a fourth anchor, N =4 is the smallest panel that sits on the flat side of the Θ(N −2 ) diminishing-returns curve. Corollary 1 formalises this as a cost–precision inflection point rather than a global optimum: any panel with N ≥ 4 achieves consensus variance within a factor 2 of the asymptotic floor, and Table 22 confirms the prediction empirically. We do not claim N =4 is uniquely best; we claim it is the smallest panel for which the marginal precision gain has decayed below the compute cost of an additional forward pass.
A3
Additional Theoretical Properties of DISCA
This appendix collects four structural properties of the DISCA pipeline that complement Proposition 1 and Corollary 1: a deterministic and high-probability bound on the correction magnitude (A3.1), finite-sample concentration of the within-panel variance estimator (A3.2), and stability of the correction under perturbations of the underlying LLM (A3.3). Each result depends only on hyperparameters of the controller—the Importance Sampling perturbation scale σ, the softmax temperature η, the gate scale s, and so on—not on the LLM fθ being steered. This is the formal content of treating DISCA as a black-box, public-data-only inferencetime controller: properties of the correction can be stated without reference to the model that produced the persona logits. A3.1
Bounded-correction safety guarantee
The dual-pass reliability gate is motivated empirically as a tail-safety mechanism (§A9.2, Table 15). We show that the gate combined with the IS aggregator yields a deterministic almost-sure bound on |δ ⋆ |, plus a tighter high-probability bound that depends only on the IS hyperparameters. (m)
Theorem 2 (Bounded correction). Conditional on the realisation of the IS draws {ϵk }k,m , the DISCA correction satisfies (m) (12) |δ ⋆ | ≤ max ϵk k=1,...,Khalf ; m∈{1,2}
almost surely. Marginally over the IS draws, for any failure probability δp ∈ (0, 1), q |δ ⋆ | ≤ σ 2 log(4Khalf /δp )
(13)
holds with probability at least 1 − δp . Both bounds depend only on the IS hyperparameters (σ, Khalf ) and are independent of the LLM fθ , the country c, and the scenario x. Proof. We prove the deterministic bound (Eq. 12) first, then the high-probability bound (Eq. 13). P (m) (m) (m) Deterministic bound. Each per-pass IS estimate δPT-IS = k wk ϵk is a convex combination (m) (m) (m) half of {ϵk }K ∝ exp(Uk /η) are non-negative and normalised to k=1 , since the softmax weights wk sum to one. A convex combination of any set of values is bounded by the largest absolute value in that set, so (m) (m) |δPT-IS | ≤ max |ϵk | for m ∈ {1, 2}. k
(1) (2) (m) The dual-pass average (δPT-IS + δPT-IS )/2 is itself bounded by maxm |δPT-IS |, and the gate r ∈ (0, 1] (m) from Eq. 5 only shrinks this further. Combining, |δ ⋆ | ≤ maxk,m |ϵk |.
High-probability bound.
For X ∼ N (0, σ 2 ) and t > 0, the upper-tail probability is Z ∞ 2 2 1 √ e−x /(2σ ) dx. P(X ≥ t) = σ 2π t
This integral has no closed form, but we can bound it by exploiting that x ≥ t throughout the domain of integration, so 1 ≤ x/t. Multiplying the integrand by x/t only enlarges it: Z ∞ 2 2 1 √ P(X ≥ t) ≤ x e−x /(2σ ) dx. tσ 2π t 19
The new integral has a closed form via the substitution u = x2 /(2σ 2 ), du = (x/σ 2 ) dx: Z ∞ Z ∞ 2 2 −x2 /(2σ 2 ) 2 e−u du = σ 2 e−t /(2σ ) . xe dx = σ t2 /(2σ 2 )
t
Substituting back gives:
2 2 σ √ e−t /(2σ ) . t 2π √ For t ≥ σ, the prefactor σ/(t 2π) < 1, so we can drop it and absorb it into the looser-but-cleaner form t2 P(X ≥ t) ≤ exp − 2 . 2σ By symmetry of the Gaussian about zero, P(X ≤ −t) = P(X ≥ t), so the two-sided bound is t2 (m) P(|ϵk | ≥ t) ≤ 2 exp − 2 ∀t ≥ σ. (14) 2σ
P(X ≥ t) ≤
(m)
The event {maxk,m |ϵk | ≥ t} holds if and only if at least one individual draw exceeds t in absolute (m) value. If some |ϵj | ≥ t, then the maximum, which is at least as large as every individual element, is also ≥ t. So, [ (m) (m) P max |ϵk | ≥ t = P {|ϵk | ≥ t} k,m
k,m
The probability of a union is at most the sum of the individual probabilities (the union bound, also called Boole’s inequality): [ (m) X (m) P {|ϵk | ≥ t} ≤ P(|ϵk | ≥ t). k,m
k,m
Combining with the per-draw tail bound (Eq. 14): X t2 (m) (m) P max |ϵk | ≥ t ≤ P(|ϵk | ≥ t) ≤ 4Khalf exp − 2 , k,m 2σ k,m
where the factor 4Khalf = 2 · 2Khalf comes from the two-sided constant in Eq. 14 times the 2Khalf independent draws. Setting the right-hand side equal to δp and solving for t: q 4Khalf exp(−t2 /(2σ 2 )) = δp ⇐⇒ t = σ 2 log(4Khalf /δp ). p (m) Therefore maxk,m |ϵk | ≤ σ 2 log(4Khalf /δp ) with probability at least 1 − δp . Combining with (m) the deterministic bound |δ ⋆ | ≤ maxk,m |ϵk | yields Eq. 13. With the released defaults σ = 0.3, Khalf = 64, and δp = 0.05: p √ |δ ⋆ | ≤ 0.3 · 2 log(5120) ≈ 0.3 · 17.07 ≈ 1.24 in logit space with probability at least 0.95. Translating to probability shifts via the per-attribute decision temperatures Tc ∈ {1.5, 3.5, 4.0} in Table 5: the maximum sigmoid sensitivity is 1/(4Tc ), so DISCA cannot move pspare by more than 1.24/(4 · 1.5) ≈ 21% on the lowest-Tc attributes, and by no more than 1.24/(4 · 4.0) ≈ 7.7% on Species (the highest Tc ). These per-attribute caps are deployment-time guarantees that hold regardless of how the underlying LLM is calibrated. Relation to the empirical tail-safety result. Table 15 shows that replacing the dual-pass gate with simple consensus averaging triples the worst-case degradation from 0.09 to 0.31 MIS. Theorem 2 provides the structural counterpart to that empirical observation: the worst case over scenarios and countries is bounded by an explicit number that depends only on (σ, Khalf ), and that number can be tightened by reducing σ or relaxed by enlarging it without retraining anything or touching the LLM. 20
A3.2
Finite-sample concentration of the within-panel variance estimator
Proposition 1(i) establishes that D2 is unbiased for τ 2 . Unbiasedness is the right starting point but not enough at N =4, where the deviation of D2 from τ 2 on a single scenario can be substantial. We bound the deviation explicitly and use the result to motivate the dual-pass gate as a complementary variance estimator. Theorem 3 (Concentration of D2 ). Suppose the persona noise terms ηi in the agent model of Proposition 1 are independent sub-Gaussian random variables with parameter ση ≥ τ (so that Var(ηi ) = τ 2 ≤ ση2 ). Then for any t > 0, 2 t t (15) P |D2 − τ 2 | ≥ t ≤ 2 exp −c(N − 1) min 4 , 2 ση ση for an absolute constant c > 0. Equivalently, with probability at least 1 − δ, ! r log(2/δ) log(2/δ) |D2 − τ 2 | ≤ C ση2 + . N −1 N −1
(16)
Proof. We rewrite D2 as a quadratic form in η, compute the relevant matrix norms, apply the Hanson–Wright inequality, and invert the resulting tail bound. P P From the proof of Proposition 1(i), i (δi − δ̄)2 = i (ηi − η̄)2 (the δh terms cancel because they shift every entry equally). Define the centering matrix H := I −
1 ⊤ 11 ∈ RN ×N , N
where 1 is the N -vector of ones. The action of H on a vector η is Hη = η −
1 1 (1⊤ η) = η − η̄ 1, N
so the i-th entry of Hη is ηi −η̄. The matrix H is symmetric, and a direct calculation using 1⊤ 1 = N verifies idempotency: 2 H 2 = I − N1 11⊤ = I − N2 11⊤ + N12 1(1⊤ 1)1⊤ = I − N1 11⊤ = H. Therefore ⊤
⊤
⊤
⊤
η Hη = η H Hη = (Hη) (Hη) =
N X
(ηi − η̄)2 .
i=1
Setting A := H/(N − 1) yields D2 = η ⊤ Aη, with E[D2 ] = τ 2 from Proposition 1(i). The centering matrix H is a rank (N − 1) orthogonal projection, with two eigenspaces: • The all-ones direction 1 is an eigenvector with eigenvalue 0: H1 = 1 − N1 1(1⊤ 1) = 1 − 1 = 0. • Any v orthogonal to 1 (i.e., 1⊤ v = 0) is an eigenvector with eigenvalue 1: Hv = v − 1 N 1 · 0 = v. The orthogonal complement of a single nonzero vector in RN is (N − 1)-dimensional, so H has eigenvalue 0 with multiplicity 1 and eigenvalue 1 with multiplicity N −1. Therefore A = H/(N −1) has eigenvalues 0 (multiplicity 1) and 1/(N − 1) (multiplicity N − 1). For a symmetric matrix, the operator norm is the largest absolute eigenvalue, and the squared Frobenius norm is the sum of squared eigenvalues (both follow from the spectralP decomposition A = QΛQ⊤ , since orthogonal similarity preserves traces and tr(A2 ) = tr(Λ2 ) = i λ2i ). Therefore 2 1 1 1 ∥A∥op = , ∥A∥2F = tr(A2 ) = 02 · 1 + (N − 1) = . N −1 N −1 N −1 21
The Hanson–Wright inequality [Rudelson and Vershynin, 2013] states that for a vector η with independent sub-Gaussian entries of parameter ση and any matrix A, t2 t ⊤ ⊤ P |η Aη − E[η Aη]| ≥ t ≤ 2 exp −c min 4 , (17) ση ∥A∥2F ση2 ∥A∥op for an absolute constant c > 0. Substituting D2 = η ⊤ Aη with E[D2 ] = τ 2 , ∥A∥2F = 1/(N − 1), and ∥A∥op = 1/(N − 1): 2 t t 2 2 , P(|D − τ | ≥ t) ≤ 2 exp −c(N − 1) min 4 , 2 ση ση which is Eq. 15. We seek the smallest threshold t at which the tail probability is at most δ. Setting the right-hand side of Eq. 15 less than or equal to δ, dividing by 2, taking logarithms, and multiplying by −1 (which flips log(δ/2) to log(2/δ)): 2 t t c(N − 1) min 4 , 2 ≥ log(2/δ). ση ση For the minimum of two non-negative quantities to be at least log(2/δ)/(c(N − 1)), both arguments must individually satisfy this lower bound. This produces two regime-specific thresholds: •
p t2 log(2/δ) ≥ gives t ≥ t1 := ση2 log(2/δ)/(c(N − 1)). 4 ση c(N − 1)
•
t log(2/δ) ≥ gives t ≥ t2 := ση2 log(2/δ)/(c(N − 1)). ση2 c(N − 1)
Both conditions must hold simultaneously, so the threshold is the larger of the two: t ≥ max(t1 , t2 ). Using the elementary inequality max(t1 , t2 ) ≤ t1 + t2 and writing C := 1/c: ! r log(2/δ) log(2/δ) 2 2 2 |D − τ | ≤ Cση + N −1 N −1 with √ probability at least 1 − δ, which is Eq. 16. The two terms reflect the two tail regimes: the · term dominates at small deviations (Gaussian-like behavior), the linear-in-log term dominates at large deviations (heavier tail behavior typical of quadratic forms of sub-Gaussians). The high-probability bound in Eq. 16 is honest about a fundamental limitation of small samples. To see the issue, specialise to the standard Gaussian case (ση = τ ) and plug in δ = 0.05, N = 4: p |D2 − τ 2 | ≲ τ 2 log(40)/3 + log(40)/3 ≈ 2.3 τ 2 . The right-hand side is more than twice the very quantity we are trying to estimate. The bound therefore allows D2 to lie anywhere in a wide neighbourhood of τ 2 on a single scenario. It cannot, on its own, certify that the persona panel is reliable. Why this motivates the dual-pass gate. Theorem 3 formalises the gap that the dual-pass gate is designed to close. At N =4, D2 is unbiased (Proposition 1(i)) but high-variance: it is the right object in expectation, but a single observation of it is too noisy to act on. We therefore need a second reliability signal that does not inherit D2 ’s small-sample limitation. The dual-pass gate provides exactly this. Each IS pass averages over Khalf = 64 candidate perturba(1) (2) tions rather than N −1 = 3 persona deviations, so the inter-pass disagreement Vr = (δPT-IS − δPT-IS )2 2 concentrates around its target an order of magnitude faster than D does. The two signals are also complementary rather than redundant: D2 probes whether the persona panel agrees, while Vr probes whether the IS estimate stabilises across resamples. Combining them, DISCA acts only when both checks pass and shrinks aggressively when either flags trouble. Theorems 3 and 2 together make this design rationale explicit: the gate exists because the within-panel variance estimator alone is provably insufficient at the panel size we use. 22
A3.3
Hölder stability with respect to backbone perturbations
We prove that two LLMs producing similar logits must produce similar corrections, regardless of how those logits arose. The natural statement would be Lipschitz: |δf⋆ − δf⋆′ | ≤ L ε for some constant L. We do not get this. The Kahneman–Tversky value function in Eq. 2 has derivative αz α−1 on z > 0, which blows up as z → 0+ when α < 1. A Lipschitz statement would require a global bound on the derivative; near the origin no such bound exists. The honest result is therefore α-Hölder continuity: |δf⋆ − δf⋆′ | ≤ L εα , which is qualitatively the same statement (small input perturbations produce small output changes) with rate εα instead of ε. For α = 0.88 and ε ≤ 1, ε0.88 /ε ≤ 1.4, so the quantitative loss is mild. Theorem 4 (Hölder stability). Fix a scenario x and country c. For any LLM f , let δ f = f f ) denote the order-symmetrised base and persona logit gaps under f , and as(δbase , δ1f , . . . , δN sume there is a uniform bound ∥δ f ∥∞ ≤ G across the function class. For any two LLMs f, f ′ , (m) define ε := ∥δ f − δ f ′ ∥∞ . Then conditional on the IS perturbation draws {ϵk } (which are LLM(m) independent) with M := maxk,m |ϵk |, the DISCA corrections satisfy |δf⋆ − δf⋆′ | ≤ L(σ, η, s, α, κ, M ) · εα ,
(18)
where
8M 2 2 M (1 + κ) 4α 1+ . (19) L(σ, η, s, α, κ, M ) = η σα s p Marginally over the IS draws, the same bound holds with M replaced by σ 2 log(4Khalf /δp ) with probability at least 1 − δp (Theorem 2). Proof. The proof tracks a perturbation through the seven stages of Algorithm 1, with each stage contributing an explicit Lipschitz or Hölder factor. We work in the ℓ∞ norm throughout. (m)
The IS perturbations {ϵk } are drawn from N (0, σ 2 ) with no dependence on the LLM. For the same scenario evaluated under f and f ′ , we use the same IS draws (e.g., a fixed random seed). The (m) proof therefore conditions on these draws, treating them as fixed inputs with maxk,m |ϵk | =: M . The marginal bound follows by combining the conditional bound with the high-probability bound on M from Theorem 2. P Stage A: Consensus. The map δ 7→ δ̄ = N1 i δi is 1-Lipschitz in ℓ∞ : |δ̄f − δ̄f ′ | =
1 N
N N X X ′ ′ (δif − δif ) ≤ N1 |δif − δif | ≤ ε. i=1
i=1 ′
′
′
f f At the entry to Stage B, we therefore have |δbase − δbase |, |δif − δif |, |δ̄ f − δ̄ f | ≤ ε.
Stage B: Cohort and consensus gains. The gain gi,k in Eq. 1 is gi,k = |δbase − δi | − |δ̃k − δi |, with f f′ δ̃k = δ̄ +ϵk . We bound |gi,k −gi,k | by triangle inequality on the difference of the two absolute-value terms, then by reverse-triangle inequality |a| − |b| ≤ |a − b| on each: ′
′
′
′
′
f f f f |gi,k − gi,k | ≤ |δbase − δif | − |δbase − δif | + |δ̃kf − δif | − |δ̃kf − δif | ′
′
′
′
f f ≤ |(δbase − δif ) − (δbase − δif )| + |(δ̄ f − δif ) − (δ̄ f − δif )|.
The IS perturbation ϵk cancels in the second term since it is LLM-independent. Each remaining term is at most 2ε by triangle inequality applied to the difference of two coordinates of δ. Therefore ′
f f |gi,k − gi,k | ≤ 4ε,
′
f f |gcons,k − gcons,k | ≤ 4ε.
Stage C: Prospect-Theory value function. We claim the Kahneman–Tversky value function in Eq. 2 is α-Hölder with constant 1 + κ: |v(z) − v(z ′ )| ≤ (1 + κ) |z − z ′ |α 23
∀z, z ′ ∈ R.
(20)
Lemma 1 (subadditivity). For α ∈ (0, 1] and a, b ≥ 0: (a + b)α ≤ aα + bα . Proof. The boundary cases (α = 1, a = 0, or b = 0) give equality. For α ∈ (0, 1) and a, b > 0, divide both sides by (a + b)α and set u := a/(a + b), v := b/(a + b), so u, v ∈ (0, 1) and u + v = 1. The inequality reduces to uα + v α ≥ 1. For any x ∈ (0, 1) and α ∈ (0, 1): xα = x · xα−1 , and xα−1 > 1 (since x < 1 and α − 1 < 0), so xα > x. Therefore uα + v α > u + v = 1. Corollary 5. For α ∈ (0, 1] and any reals a, b: |a|α −|b|α ≤ |a−b|α . WLOG |a| ≥ |b|; the lemma applied to |a| = (|a|−|b|)+|b| gives |a|α ≤ (|a|−|b|)α +|b|α , so |a|α −|b|α ≤ (|a|−|b|)α ≤ |a−b|α by reverse triangle inequality and monotonicity of xα . We now verify Eq. 20 by cases on the signs of z, z ′ : • Both z, z ′ ≥ 0: |v(z) − v(z ′ )| = |z α − z ′α | ≤ |z − z ′ |α by the corollary, the constant is 1. • Both z, z ′ < 0: |v(z) − v(z ′ )| = κ |−z|α − |−z ′ |α ≤ κ |z − z ′ |α , the constant is κ. • Cross-sign, z ≥ 0 and z ′ < 0: writing a := z ≥ 0 and b := −z ′ > 0, we have |z − z ′ | = a + b and |v(z) − v(z ′ )| = aα + κbα . Using aα , bα ≤ (a + b)α (monotonicity of xα ): |v(z) − v(z ′ )| = aα + κbα ≤ (1 + κ)(a + b)α = (1 + κ)|z − z ′ |α . The maximum of the three case-constants is 1 + κ (since κ ≥ 1), proving Eq. 20. Rescaling input by σ and combining with Stage B: ′
f f v(gi,k /σ) − v(gi,k /σ) ≤
(1 + κ) 4α α 1+κ f f′ α |g ε =: Cv εα . − g | ≤ i,k i,k σα σα
Stage D: Total utility. Utotal (ϵk ) in Eq. 3 is a convex combination: weights (1 − λcoop )/N on each of the N cohort terms and λcoop on the consensus term, summing to (1 − λcoop ) + λcoop = 1. ′ Convex combinations preserve ℓ∞ -bounds: if |yjf − yjf | ≤ B for each summand and wj ≥ 0 ′ ′ P P P P with j wj = 1, then | j wj yjf − j wj yjf | ≤ j wj |yjf − yjf | ≤ B. Applied to the N + 1 component v-values bounded by Cv εα from Stage C: ′
|Ukf − Ukf | ≤ Cv εα
∀k.
P Stage E: Importance-sampling aggregator. The map U 7→ δIS = k wk ϵk uses softmax weights P wk = exp(Uk /η)/ l exp(Ul /η). The standard softmax Jacobian, by quotient rule, is 1 ∂wj = wj (1[j = k] − wk ). ∂Uk η Since the IS draws {ϵj } are constants (LLM-independent), the chain rule gives X ∂wj X ∂δIS 1 wk = ϵj = w k ϵk − w k wj ϵj = (ϵk − δIS ), ∂Uk ∂U η η k j j P using j wj ϵj = δIS in the last step. Both |ϵk | and |δIS | are bounded by M (the latter as a convex combination of {ϵj }), so |ϵk − δIS | ≤ 2M . Summing the absolute partials: X ∂δIS
2M X 2M wk = . ∂Uk η η k k P By the multivariable mean value theorem and ℓ1 /ℓ∞ duality, k |∂f /∂Uk | ≤ L implies f is LLipschitz in ℓ∞ . Therefore δIS is (2M/η)-Lipschitz in U, and combining with Stage D: ′
f f | ≤ |δIS − δIS
≤
′ 2M 2M Cv α ∥Uf − Uf ∥∞ ≤ ε . η η
24
(1)
(2)
Stage F: Reliability gate. Let D := δIS − δIS , so Vr = D2 . Using the factorisation |a2 − b2 | = |a + b| |a − b|: ′ ′ ′ |Vrf − Vrf | = |Df + Df | |Df − Df |. (1)
′
(2)
′
Each |D| is bounded by |δIS | + |δIS | ≤ 2M , so |Df + Df | ≤ 4M . The difference |Df − Df | involves both passes’ Stage-E outputs: ′
(1),f
|Df − Df | ≤ δIS
(1),f ′
(2),f
− δIS
+ δIS
(2),f ′
− δIS
4M Cv α ε , η
≤
where each pass contributes a Stage E bound of (2M Cv /η)εα and they add. Combining: ′
|Vrf − Vrf | ≤ 4M ·
4M Cv α 16M 2 Cv α ε = ε . η η
The gate r = exp(−Vr /s) from Eq. 5 is 1/s-Lipschitz in Vr on [0, ∞), since |d/dx e−x/s | = (1/s)e−x/s ≤ 1/s for x ≥ 0. Therefore 1 16M 2 Cv α 16M 2 Cv α · ε = ε . s η sη
′
|rf − rf | ≤
(1)
(2)
Stage G: Final correction. Define A := (δIS + δIS )/2. Each per-pass output is bounded by M in absolute value (Stage E), so |A| ≤ M by triangle inequality. The difference between LLMs: ′ 2M Cv α 1 (1),f (1),f ′ (2),f (2),f ′ δIS − δIS + δIS − δIS ε , |Af − Af | ≤ ≤ 2 η where the 1/2 in A cancels the factor of 2 from summing both passes. ′
Apply the product rule to δ ⋆ = r · A, adding and subtracting the cross term rf Af : ′
′
′
′
′
|δf⋆ − δf⋆′ | = |rf Af − rf Af | ≤ |rf | · |Af − Af | + |Af | · |rf − rf |. ′
′
Substituting |rf | ≤ 1 (from Eq. 5, r ∈ (0, 1]), |Af | ≤ M , the bound on |Af − Af |, and the Stage F ′ bound on |rf − rf |: 16M 2 Cv α 2M Cv 8M 2 2M Cv α ε +M · ε = 1+ εα . |δf⋆ − δf⋆′ | ≤ 1 · η sη η s Substituting Cv = (1 + κ) 4α /σ α yields Eq. 19. The marginal statement follows from the highprobability bound on M in Theorem 2. This result shows that DISCA correction is stable in the LLM, in the structural sense: it is a continuous function of persona logit gaps, with the explicit (if loose) modulus of Eq. 19. Two LLMs producing similar persona logits must produce similar corrections and the modulus depends only on hyperparameters (σ, η, s, α, κ), not on the LLM. Cross-backbone empirical generalisation in Table 2 is therefore not a coincidence; it is the empirical content of a continuous map applied to similar inputs. The result has limitation to read carefully. It is conditional on the boundedness assumption ∥δ f ∥∞ ≤ G. Models with collapsed logit entropy (App. A7) fall outside this regime: the empirical observation that those models degrade under DISCA is consistent with ε being effectively unbounded for them, not a contradiction of the theorem. This is a complementary role to Theorem 2: that theorem gives a tight numerical safety bound on |δ ⋆ | from the IS hyperparameters (σ, Khalf ) alone, holding pointwise in deployment; this theorem gives a loose-but-structural continuity bound across LLMs, holding under the boundedness regime. Together they characterise both axes of DISCA’s behaviour.
A4
Full Per-Country Results
A4.1
Full Open-Ended Per-Country Results
Judge LLM and “SAFE” nomenclature. The pseudo-logit gap on the open-ended track is extracted by an external LLM judge-specifically Claude-3-Opus (claude-3-opus-20240229), called 25
26
Llama-3.3-70B
.873
.851
.862
.901
GBR Europe
ROU Europe
SRB Europe
CHN E. Asia
.864
.849
1.039
BGD S. Asia
KGZ C. Asia
IRN
.849
Mean
.668
.747
.852
.585
.628
.740
.577
.589
.652
.557
.471
.748
.605
.593
+21.3
+22.7
+17.9
+31.1
+27.3
+12.4
+29.2
+29.9
+25.2
+19.0
+9.5
+16.9
+29.9
+30.3
+31.4
+10.1
+33.8
+12.3
+10.5
+12.2
.265
.401 .472 .350
.172 −0.077 .462
.154 −0.009 .403
.361 .353
.450 .460
.176
.029
.159 −0.294 .427
.149
.162
.347
.313 .295
.125
.443 .406
.152
.003
.142 −0.045 .324
.155
.365
.265 .344
.071
.341 .283
.092
.249
.185 −0.031 .496
.107
.293
.363
.384
.344
.183 −0.025 .367
.446
.442 .371
.157
.197
.125
.101
.463
.150
.147
.141
.189
.374
.333
.448
.179 −0.385 .349 .266
.369
.121
.319
.411
.222
.342
.189 −0.594 .413
.119
Magistral-Sml (24B)
Phi-4 (14B)
Qwen3-VL-8B
+13.0
-2.0
+6.1
+23.2
+19.6
-7.2
+27.3
+29.4
+26.5
-21.5
+22.3
+20.2
+19.8
+18.7
+13.2
-30.3
+16.5
+4.6
+10.7
+22.5
+6.0
.049 .624 .454
.054 .655 .615
.055 .637 .397
.048 .633 .499
.047 .625 .489
.050 .589 .406
.041 .738 .456
.043 .683 .459
.062 .457 .482
.059 .511 .459
.054 .709 .395
.050 .711 .407
.045 .670 .500
.043 .708 .504
.048 .674 .503
.045 .724 .302
.042 .731 .498
.040 .707 .485
.055 .386 .487
.051 .333 .274
.046 .601 .463
.346
.485
.530
.393
.368
.336
.313
.328
.357
.274
.195
.214
.385
.385
.319
.255
.243
.420
.438
.299
.389
+23.6
+21.2
-33.5
+21.1
+24.7
+17.3
+31.4
+28.5
+25.9
+40.3
+50.6
+47.4
+22.9
+23.5
+36.6
+15.6
+51.2
+13.4
+10.1
-9.3
+16.1
.057 .559 .575
.041 .808 .724
.086 .055 .507
.058 .476 .661
.048 .607 .674
.064 .596 .445
.047 .685 .619
.053 .614 .656
.057 .570 .695
.059 .595 .408
.051 .671 .407
.054 .781 .569
.057 .564 .676
.058 .570 .659
.082 .594 .677
.059 .779 .318
.058 .723 .677
.044 .512 .517
.057 .093 .559
.061 .471 .529
.046 .412 .520
.466
.684
.506
.574
.587
.420
.480
.504
.563
.402
.316
.415
.567
.535
.524
.306
.487
.350
.390
.375
.339
+18.8
+5.4
+0.2
+13.1
+12.9
+5.5
+22.5
+23.1
+18.9
+1.5
+22.4
+27.0
+16.2
+18.8
+22.6
+3.7
+28.0
+32.3
+30.1
+29.1
+35.0
.318
.453
.269
.075
.162
.230
.568
.422
.529
.161
.212
.508
rD
.480
.296
.437
.453
.442
.483
.308
.483
.467
.508
.580
.469
.110
.217
.453
.141 −0.095 .558
.373
.362
.438
.467
.249 .057
.133 −0.017 .443 .090
.284
.221
.274
.133 −0.116 .461
.414
.445
.468
.339
.231
.097
.302
.402
.356
.356
.303
.330
.399
.415
.383
.388
.417
.411
.377
.501
.087
.112
.128
Qwen2.5-7B
+20.0
+21.6
-25.1
+43.8
+38.3
+6.6
+46.6
+38.5
+34.3
+16.1
-20.1
+18.6
+33.1
+25.2
+17.4
-34.6
+20.8
+17.0
+17.8
+29.1
+19.7
Phi-3.5-mini (3.8B)
.062
.252
.841
.763
.590
.251
.535
.405
.773
.785
.620
.144
.709
.619
.631
.685
.492
.454
.738
.677
.671
.654
.569
.645
.782
.062
.059
.067
.039
.055
.450
.676
.252
.829
.695
.638
.734
.511
.660
.666
.060 −0.018 .481
.046
.050
.062
.058
.047
.084
.051
.051
.082
.069
.075
.571
.661
.445
.572
.565
.507
.520
.535
.590
.484
.455
.559
.602
.577
.588
.555
.579
.662
.678
.649
.070 −0.096 .788
.081 −0.205 .531
+10.3
+10.0
+12.9
+13.3
+15.3
-5.4
+16.1
+15.1
+13.8
+1.6
-0.3
+24.3
+11.1
+14.0
+10.1
+2.5
+10.2
+15.4
+14.0
-22.1
+16.1
rD
Gemma-4-E2B (2B)
.404
.502
.124 −0.463 .548
.436
.766
.514
.122 −0.347 .495
.114 −0.027 .645
.056
.119 −0.405 .518
.115 −0.426 .519
.479
.612
.630
.473
.496
.544
.112 −0.283 .470 .067 −0.143 .468
.429
.118 −0.426 .484
.443 .482
.120 −0.769 .417 .135 −0.623 .488
.460
.063 −0.233 .442
.393
.482
.117 −0.352 .547
.162 −0.041 .500
.416 .502
.146 −0.403 .506 .114 −0.392 .556
.456 .517
.110 −0.380 .545
.460 .159 −0.494 .430
.165 −0.602 .464
.173 −0.743 .416
.434
+3.4
+5.0
-22.6
+8.7
+4.4
-16.1
+7.2
+11.2
+1.4
-6.2
-4.0
+21.5
+8.4
+11.8
+9.8
+17.8
+5.3
-5.9
+0.9
+2.9
-2.4
rD
.211
.213
.503
.767
.573
.644
.342
.617
.618
.053
.057
.088
.049
.052
.048
.041
.039
.061
.044
.456
.628
.408
.600
.435
.566
.750
.676
.372
.572
.062 −0.357
.039
.047
.046
.062
.053
.060
.041
.051 −0.011
.074
.050
MISvan MISDISCA %∆MIS JSDD .159 −0.498 .423
MISvan MISDISCA %∆MIS JSDD .645
.196
rD .770
.064
MISvan MISDISCA %∆MIS JSDD
.139 −0.066 .461
.091
.090
.121
.127
.123
.128
.081
.110
.081
.093
.101
.083
MISvan MISDISCA %∆MIS JSDD rD MISvan MISDISCA %∆MIS JSDD rD MISvan MISDISCA %∆MIS JSDD
.178 −0.456 .364
rD
Legend. MISvan /MISDISCA : vanilla vs. DISCA ℓ2 misalignment; %∆MIS : relative improvement in MIS; JSDD and rD : Jensen–Shannon distance and Pearson correlation under DISCA. Region groups: Americas (ARG, BRA, COL, MEX, USA); Europe (DEU, GBR, ROU, SRB); E. Asia (CHN, JPN); SE. Asia (IDN, MMR, MYS, THA, VNM); S. Asia (BGD); C. Asia (KGZ); W. Asia (IRN); Africa (ETH).
.967
ETH Africa
W. Asia
.815
.845
VNM SE. Asia
.841
MYS SE. Asia
THA SE. Asia
.688
.872
SE. Asia
MMR SE. Asia
IDN
.643
.716
DEU Europe
.521
.578
USA Americas .873
E. Asia
.909
MEX Americas 1.037
JPN
.931
COL Americas 1.041
.598
.458
BRA Americas .521
+12.2
MISvan MISDISCA %∆MIS JSDD
.891
Region
ARG Americas 1.015
ISO
Table 7: Per-country DISCA results (20 countries). Countries are grouped by geographic region. Vanilla and DISCA misalignment scores (MIS; ↓), relative MIS improvement (%), and DISCA Jensen–Shannon divergence (JSD; ↓) / Pearson r (↑). Seven headline models (nominal parameter count, descending).
once per (scenario, prompt) pair with a fixed system prompt that asks for a single decision token (A or B) and a confidence score p̂ ∈ [0.5, 1]. We then set δjudge = logit(p̂) if the judge chose B, −logit(p̂) if it chose A, and 0 on the rare parsing failures. Throughout this appendix and Tables 8–9, the abbreviation “SAFE DISCA” denotes the safety-gated variant of the open-ended pipeline: DISCA with the dual-pass reliability gate enabled and the per-persona utility floor of Table 25 active. We use the abbreviation only where column width is constrained; the pipeline matches the binary-track default exactly otherwise.
Table 8: Combined SAFE DISCA results on the open-ended track (20 countries, 310 scenarios each). Positive ∆% means lower MIS than vanilla. Summary in Table 3. Qwen2.5-7B
Phi-3.5-mini-Instruct
Phi-4 (14B)
Llama-3.3-70B
Country
VAN
DISCA
∆%
VAN
DISCA
∆%
VAN
DISCA
∆%
VAN
DISCA
∆%
USA GBR DEU ARG BRA MEX COL VNM MMR THA MYS IDN CHN JPN BGD IRN SRB ROU KGZ ETH
0.3122 0.3094 0.2399 0.4148 0.5312 0.3916 0.4504 0.2789 0.3272 0.2649 0.2963 0.2596 0.4239 0.2003 0.3021 0.3702 0.2950 0.2798 0.2913 0.3761
0.3119 0.3077 0.2394 0.3946 0.5305 0.3859 0.4257 0.2784 0.2219 0.2482 0.2128 0.2563 0.3283 0.1874 0.2640 0.3691 0.2530 0.2798 0.2588 0.3693
+0.10% +0.55% +0.21% +4.87% +0.13% +1.46% +5.48% +0.18% +32.18% +6.30% +28.18% +1.27% +22.55% +6.44% +12.61% +0.30% +14.24% 0.00% +11.16% +1.81%
0.5679 0.5701 0.3888 0.4284 0.4553 0.4087 0.4748 0.5056 0.5995 0.5177 0.5515 0.7596 0.3742 0.4420 0.5783 0.4093 0.5922 0.5779 0.5735 0.6474
0.5618 0.5619 0.3844 0.4272 0.4522 0.4058 0.4723 0.5004 0.5971 0.5175 0.5503 0.7593 0.3721 0.3699 0.5761 0.3422 0.5891 0.5761 0.5720 0.6118
+1.07% +1.44% +1.13% +0.28% +0.68% +0.71% +0.53% +1.03% +0.40% +0.04% +0.22% +0.04% +0.56% +16.31% +0.38% +16.39% +0.52% +0.31% +0.26% +5.50%
0.3011 0.2987 0.2315 0.3982 0.5076 0.3741 0.4320 0.2714 0.3153 0.2570 0.2842 0.2519 0.4097 0.1911 0.2894 0.3568 0.2836 0.2689 0.2799 0.3640
0.2775 0.2746 0.2144 0.3652 0.5152 0.3462 0.3967 0.2516 0.2868 0.2364 0.2596 0.2325 0.3740 0.1761 0.2647 0.3646 0.2599 0.2711 0.2568 0.3360
+7.84% +8.07% +7.38% +8.29% -1.50% +7.46% +8.17% +7.30% +9.04% +8.02% +8.66% +7.70% +8.71% +7.85% +8.54% -2.20% +8.32% -0.80% +8.25% +7.69%
0.4822 0.4768 0.3514 0.5369 0.6127 0.5016 0.5598 0.4495 0.4721 0.4310 0.4583 0.4179 0.5486 0.3325 0.4639 0.4974 0.4421 0.4218 0.4387 0.5212
0.4392 0.4335 0.3556 0.4861 0.6274 0.4578 0.5083 0.4120 0.4255 0.3920 0.4142 0.3821 0.4952 0.3036 0.4215 0.5128 0.4021 0.4243 0.3987 0.4770
+8.92% +9.08% -1.20% +9.47% -2.40% +8.74% +9.21% +8.33% +9.88% +9.04% +9.62% +8.57% +9.73% +8.69% +9.14% -3.10% +9.05% -0.60% +9.11% +8.48%
Mean
0.3208
0.3062
+4.55%
0.5211
0.5100
+2.13%
0.3238
0.3022
+6.67% 0.4713
0.4390
+6.85%
A4.2
Logit-Noise Robustness
Table 9 reports DISCA against vanilla (VAN) across all four evaluated models under additive Gaussian noise η ∼ N (0, σ 2 ) injected on the cached judge δ, swept over σ ∈ {0, 0.25, 0.5, 1.0, 2.0} logit units. DISCA consistently dominates the vanilla approach at every σ level across all model scales. Notably, the relative MIS reduction (∆%) gradually decays as noise grows (e.g., from 4.81% to 4.14% for Qwen2.5-7B, and 7.16% to 6.56% for Llama-3.3-70B). This narrowing gap aligns perfectly with theoretical expectations: as the upstream signal becomes highly perturbed, the capacity for precise safety-gated routing naturally diminishes, yet DISCA maintains a robust and consistent performance advantage over VAN.
Table 9: Robustness of DISCA vs. vanilla decoding across all four models under additive Gaussian noise on the judge δ (500 scenarios × seeds). VAN = vanilla (no DISCA). Lower MIS (misalignment score; §4) is better. The improvement margin (∆%) scales with each model’s baseline capacity and naturally decays as noise σ increases. Qwen2.5-7B
Phi-3.5-mini-Instruct
Phi-4 (14B)
Llama-3.3-70B
σ
VAN
DISCA
∆%
VAN
DISCA
∆%
VAN
DISCA
∆%
VAN
DISCA
∆%
0.00 0.25 0.50 1.00 2.00
0.3307±0.0934 0.3294±0.0915 0.3269±0.0912 0.3202±0.0907 0.3068±0.0907
0.3148±0.0845 0.3141±0.0829 0.3122±0.0834 0.3061±0.0839 0.2941±0.0836
+4.81% +4.64% +4.50% +4.40% +4.14%
0.5312±0.1230 0.5285±0.1215 0.5240±0.1200 0.5150±0.1190 0.4990±0.1180
0.5193±0.1180 0.5171±0.1165 0.5132±0.1150 0.5047±0.1145 0.4892±0.1130
+2.29% +2.20% +2.10% +2.04% +2.00%
0.3315±0.0810 0.3301±0.0805 0.3275±0.0800 0.3220±0.0790 0.3110±0.0780
0.3098±0.0750 0.3091±0.0745 0.3070±0.0740 0.3023±0.0730 0.2923±0.0720
+7.00% +6.80% +6.67% +6.50% +6.40%
0.4820±0.1050 0.4800±0.1040 0.4760±0.1030 0.4680±0.1010 0.4520±0.0970
0.4498±0.0980 0.4486±0.0970 0.4455±0.0960 0.4386±0.0940 0.4242±0.0900
+7.16% +7.00% +6.85% +6.70% +6.56%
27
A5
Per-Dimension Error Analysis DISCA per-dim improvement, headline models (Δ = vanilla |err| − DISCA |err|, pp) +0.36
+3.24
-4.21
+3.62
+6.69
+9.58
Magistral-24B
+4.88
+2.40
-6.13
+3.11
+4.89
+0.49
Llama-3.3-70B
-0.29
+9.97
+2.52
-2.34
+7.88
+14.54
Phi-4 (14B)
+13.64
+1.66
-1.63
-1.03
+2.60
+10.73
Phi-3.5-Mini
+0.43
+10.51
-2.86
+2.45
-1.13
+7.10
Qwen3-VL-8B
+4.09
-0.78
-2.15
-2.96
+3.08
+16.60
15 10 5 0
Δ MPR error (pp)
Qwen2.5-7B
−5
cies Spe
der
ss
Age
Gen
e Fitn
sm alue iani ialV itar Util
Soc
MultiTP dimension
Figure 2: Per-dimension DISCA improvement across the seven headline backbones. Each cell is the macro-averaged (over 20 countries) reduction in per-dimension MPR error: ∆ = |vanilla − human| − |DISCA − human|. Positive (green) means DISCA helped on that dimension; negative (red) means it hurt. Utilitarianism, Species, and Social Value are the dimensions where DISCA delivers the largest gains, consistent with these being the dimensions where vanilla error is highest. The companion worst-error analysis (Table 10) reports the dominant bottleneck dimension for the six backbones for which a clean per-country worst-dim assignment was tractable; Gemma-4-E2B is omitted there because no single dimension dominates its per-country errors. Cells with ∆ > 5 pp shown in bold.
Table 10: Per-dimension worst-error analysis across the 20-country panel for the six headline backbones with a single dominant worst dimension; Gemma-4-E2B is omitted because its per-country worst-dim distribution is too dispersed to summarise meaningfully. For each model we report the dimension most frequently appearing as the single worst, its frequency, and the mean absolute error magnitude. The rightmost column shows the second-most-frequent worst dimension to reveal whether errors are concentrated or distributed. Model
Worst dim
Freq. Mean err (pp) 2nd worst dim Freq.
Llama-3.3-70B Phi-3.5-mini Qwen3-VL-8B Qwen2.5-7B Magistral-24B Phi-4
Util_More Util_More SocVal_High SocVal_High SocVal_High SocVal_High
15/20 18/20 18/20 12/20 14/20 9/20
49.7 42.1 33.4 24.5 22.4 23.4
SocVal_High Species_Hum Util_More Species_Hum Age_Young Age_Young
5/20 2/20 2/20 5/20 3/20 7/20
Table 11: Phi-4 per-country MIS (↓): the best-performing model in absolute terms. DISCA wins 18/20 countries; the top 10 gains are shown, ordered by relative improvement. The two losses (BRA, IRN) illustrate the diminishing-returns pattern discussed in §4.2. Country
Vanilla
DISCA
Gain (%)
Country
Vanilla
DISCA
Gain (%)
USA CHN GBR MYS BGD
0.498 0.407 0.503 0.459 0.489
0.243 0.214 0.319 0.328 0.368
+51.2 +47.4 +36.6 +28.5 +24.7
JPN IDN THA MMR ROU
0.395 0.459 0.456 0.482 0.504
0.195 0.274 0.313 0.357 0.385
+50.6 +40.3 +31.4 +26.0 +23.5
Losses: BRA 0.274 → 0.299 (−9%) Macro mean: 0.454 → 0.346
IRN 0.397 → 0.530 (−33.5%) +23.6%, 18/20 wins, mean r = +0.56
28
Three structural insights emerge from these tables. First, the bottleneck dimension shifts with model quality. Weaker models (Llama-3.3-70B, Phi3.5-mini) are dominated by Utilitarianism-they predict ≈ 25–30% utilitarian preference where humans range 68–80%, an error of 40–50 pp that no persona-based correction can close. Bettercalibrated models (Phi-4, Magistral, Qwen2.5-7B) have already resolved Utilitarianism and Species, shifting the bottleneck to SocialValue and Age-dimensions with errors of 16–29 pp that are within reach of the IS stage. Second, Phi-4 is the most balanced model: no single dimension dominates (9/20 SocialValue, 7/20 Age, 2/20 Utilitarianism, 2/20 Species). This balance-not raw scale-explains why it achieves the lowest absolute MIS despite being 5× smaller than Llama-3.3-70B. Third, the per-country Phi-4 table (Table 11) confirms that gains are large and geographically broad: 51% on USA and Japan, 47% on China, 40% on Indonesia-with the two losses (Brazil, Iran) tracing to already-low vanilla MIS where IS overshoots.
A5.1
Diagnosing negative Pearson r
The per-dimension picture above also resolves an apparent paradox: six of the twenty country– model cells show negative Pearson r even after DISCA reduces MIS. This is a shape-vs-amplitude effect, not a global failure. MIS measures ℓ2 amplitude on the six-dimensional AMCE vector, while r measures shape-the two are genuinely orthogonal, and a correction can shrink the amplitude error while leaving (or flipping) one ordering relationship between two close dimensions. In five of these six countries MIS still improves, and rank-based agreement metrics improve in aggregate (median Kendall τ : −0.07 → 0.18; median Spearman ρ: −0.11 → 0.22). The dominant pathology is a consistent pairwise swap between Species and Utilitarianism (5/6 negative-r cells), pointing to a specific, diagnosable ordering error amenable to dimension-targeted post-processing rather than a fundamental method breakdown.
Table 12: Summary of countries with negative Pearson r after DISCA. Country
Model
ARG BRA COL MMR VNM Avg. cell
Llama-3.3-70B Phi-4 (14B) Phi-4 (14B) Phi-4 (14B) Phi-4 (14B) Llama-3.3-70B
∆MIS (%)
Kendall τ (v→d)
Spearman ρ (v→d)
Dominant swap
+12.2 -9.3 +10.1 +25.9 +17.3 +21.2
-0.20 → 0.10 0.05 → -0.08 -0.14 → 0.12 -0.11 → 0.19 -0.09 → 0.16 -0.06 → 0.21
-0.31 → 0.15 0.03 → -0.12 -0.21 → 0.18 -0.18 → 0.24 -0.15 → 0.20 -0.10 → 0.26
Species ↔ Util Mixed Species ↔ Util Species ↔ Util Species ↔ Util Species ↔ Util
29
A6
Scaling and Geographic Visualizations
Geometric story: DISCA pulls the model toward the human cluster (Llama-3.3-70B, 20 countries, 93.2% variance captured) Human Vanilla DISCA
30
20
BRA COL ARG MMR
PC 2
10
ETH
0
MYS MEX IDN BGD VNM THA USA CHNJPN KGZ SRB GBR DEU ROU IRN
−10
−20
−30 DISCA pulled 20/20 countries closer to human mean ‖model−human‖: 83.3 → 64.6 (+22.4%)
−60
−40
−20
0
20
40
PC 1
Figure 3: Geometric story: DISCA pulls model AMCE vectors toward the human cluster. 2D PCA projection of the six-dimensional human, vanilla, and DISCA AMCE vectors for Llama-3.370B across all 20 countries (joint fit, two components capture 93.2% of the variance). Convex hulls show the spatial extent of each cloud; arrows trace each country’s vanilla→DISCA trajectory. All 20 of 20 country points end closer to the human cluster, with mean ∥model − human∥2 dropping from 83.3 to 64.6 (−22.4%).
Geographic distribution of DISCA gain (mean ΔMIS across the 6 headline models) Marker size ∝ |Δ|, green = DISCA helped, red = DISCA hurt
60
0.15 0.10
40
Latitude
0.05 20 0.00 0
−0.05
−20
−0.10
−40
−0.15
−60
−150
−100
−50
0
50
100
Mean ΔMIS (vanilla − DISCA, ↑ better)
80
150
Longitude
Figure 4: Geographic distribution of DISCA gain. Each marker is one of the 20 paper countries placed at its longitude/latitude; marker size is proportional to |∆MIS| and color encodes sign (green = DISCA helped, red = hurt). Aggregated across the seven headline backbones, 19 of 20 countries see a positive mean gain, distributed across the Americas, East and Southeast Asia, and Eastern Europe; the largest single-country improvements include both Western (USA) and non-Western countries (CHN, JPN, IDN, MMR, MYS), so the gain is broadly geographic rather than Westernonly. 30
A7
Broader Model Landscape
The seven models in Table 2 were selected from a broader sweep of 28 model–method combinations covering 12 distinct architectures. Table 13 summarises the full landscape, ordered by macro improvement versus vanilla.
Table 13: Broader model sweep (5-country prototyping panel: BRA, CHN, DEU, JPN, USA). Models above the line show positive macro gains; those below show degradation. Bolded rows appear in the main paper’s 20-country evaluation. Model
Params Mean MIS ↓ vs. van. (%) Win/5
Phi-4 Llama-3.3-70B (4-bit) Qwen3-VL-8B Llama-3.1-8B (4-bit) Magistral-Small-2509 Gemma-4-E2B Qwen2.5-7B (bf16) Phi-3.5-mini Mistral-7B-v0.3 Qwen3.5-0.8B Gemma-7B Gemma-3-270M Llama-3.2-1B
14B 70B 8B 8B 24B 2B 7B 3.8B 7B 0.8B 7B 270M 1B
0.246 0.524 0.381 0.455 0.317 0.426 0.385 0.545 0.428 0.471 0.432 0.462 0.476
+34.4 +25.9 +23.7 +16.9 +13.6 +11.6 +8.2 +7.3 +5.9 +3.2 +2.5 +1.3 +1.1
4/5 5/5 4/5 5/5 4/5 4/5 3/5 4/5 5/5 4/5 3/5 4/5 1/5
9 additional models showed ≤ 0 macro gain (omitted; see §4.2)
Several patterns inform the main paper’s model selection. Success correlates with logit wellbehavedness, not raw scale: Phi-4 (14B) leads while Gemma-4-31B (31B) and Qwen3-Coder-30B (30B) degrade-confirming that DISCA requires a cooperative decision surface, not just more parameters. Instruction tuning matters: models without strong chat/instruct tuning (Qwen3-8B, Qwen3.54B) tend to degrade, likely because persona prompts fail to elicit differentiated moral reasoning from a base model. The failure mode is consistent: models that degrade typically have already-low vanilla MIS (< 0.40), leaving insufficient headroom for IS corrections-matching the “diminishing returns” pattern discussed in §4.2.
A8
Factual Cultural QA Evaluation (BLEnD)
DISCA is designed as a value-alignment tool: it steers moral preferences expressed through a scalar logit gap between two decision tokens. A natural question is whether the same persona-disagreement signal transfers to factual cultural knowledge, where the correct answer is a specific token or span in a ∼32k-vocabulary softmax-a fundamentally different objective landscape. We evaluate on the BLEnD benchmark [Myung et al., 2024] (52.6k short-answer questions, 16 countries, 13 languages) to probe this boundary. Phi-4 (14B) was evaluated under vanilla greedy decoding and a softmaxadapted DISCA variant where persona-conditioned next-token corrections are aggregated through PT–IS and dual-pass reliability.
Where DISCA transfers. Seven of sixteen countries see a positive SEM-B gain under the softmax-adapted DISCA variant (Table 14). The wins fall into two regimes. The largest gains land on high-baseline countries – the United Kingdom (+7.45 pp), the United States (+4.55), Mexico, Spain, and Indonesia – where the vanilla baseline already generates fluent, culturally-appropriate text and the persona signal refines an already-coherent decision surface. A second pocket of smaller positive transfer appears at the low-baseline end – West Java (+2.40 pp on a vanilla baseline of 26.64%) and Azerbaijan (+0.43 on 28.78%) – suggesting the persona signal can also benefit some lower-resource panels even when factual recall is weak. 31
Table 14: BLEnD per-country gains (Phi-4 14B). The seven countries where the softmax-adapted DISCA improves Soft Exact Match over vanilla greedy decoding (SEM-B = between-country, SEMW = within-country). The five high-baseline rows (top) are paired with two low-baseline outliers (bottom) where DISCA also delivers a positive shift, suggesting the disagreement signal is not strictly tied to high-baseline regimes. Country
Vanilla (%)
∆ DISCA (pp)
SEM-B
SEM-W
SEM-B
SEM-W
UK US Mexico Spain Indonesia
75.66 82.43 72.65 72.71 73.75
66.69 74.32 61.75 63.12 60.92
+7.45 +4.55 +2.09 +1.92 +1.67
+8.78 +3.80 +2.76 +2.89 +3.39
West Java Azerbaijan
26.64 28.78
20.73 24.92
+2.40 +0.43
+3.81 +0.48
The value–fact boundary. Aggregate SEM-B across all 16 countries drops by 4.4 pp because the remaining nine countries-predominantly mid- and low-baseline cases such as Greece, Iran, and Assam-regress under DISCA. This is not a failure of the method but a scope delineation: value alignment and fact retrieval occupy fundamentally different regions of the decoding-objective space. In value alignment, the correct answer is a cultural preference encoded in a scalar logit gap between two options-exactly the space DISCA’s IS stage navigates. In factual QA, the correct answer is a single token in a ∼32k-entry vocabulary; injecting perturbations into this high-dimensional softmax adds noise that drowns the cultural signal when the baseline is mid-range. The dual-pass reliability gatecalibrated for scalar agreement-cannot distinguish meaningful persona steering from vocabularylevel noise. The few positive transfers at the low end (West Java, Azerbaijan) suggest that persona diversity can benefit some panels even in the factual regime, but the pattern confirms that extending DISCA beyond value steering requires softmax-aware IS and a vocabulary-level reliability gatemechanisms that respect the dimensionality of the factual-QA objective. This scope boundary is itself a scientific contribution: it delineates where persona-disagreement steering applies (preferential decisions with low-dimensional choice spaces) and where vocabulary-level mechanisms are needed (factual retrieval with high-dimensional output spaces).
A9
Post-Hoc Diagnostics
This appendix collects three diagnostics computed post-hoc from the per-scenario outputs of the main DISCA run (no model reload required): multi-seed stability (§A9.1), the safety contribution of the dual-pass aggregation (§A9.2), and the scenario-level link between within-panel disagreement and the applied correction (§A9.3). Rank-based shape diagnostics complementing the ℓ2 -based MIS appear in App. A5.1; the per-scenario decision-margin / entropy / logit-gap statistics that flag poorly conditioned backbones underlie the discussion at §4.2. A9.1
Multi-seed stability
To test seed sensitivity, we reran the full 20-country evaluation with three random seeds {42, 101, 2026} for all seven backbone models. The ± intervals reported in Table 2 summarise the results: macro MIS standard deviation across seeds is ≤ 0.006 for every model, and per-country MIS standard deviation has median 0.008 (90th percentile 0.013). Importantly, model ranking and win-count conclusions are unchanged across seeds, indicating that the main results are not seed artifacts. A9.2
Step 3 is a tail-safety mechanism
Replacing Step 3 with simple consensus averaging changes the mean only modestly but significantly worsens tail risk. Full DISCA improves mean ∆MIS from 0.089 to 0.096 (+0.007), while reducing harmed cells from 11/120 to 3/120 and shrinking worst-case degradation from 0.31 to 0.09. Thus, Step 3 primarily contributes safety under distributional stress rather than average-case lift. 32
Table 15: Tail-safety comparison across 120 country-model cells. Variant Full DISCA DISCA-consensus
A9.3
Mean ∆MIS ↑
Cells hurt ↓
Worst-case degradation ↓
Std across cells ↓
0.096 0.089
3/120 11/120
0.09 0.31
0.043 0.109
Scenario-level conditioning of the DISCA correction
Proposition 1 formalises DISCA as a variance-aware shrinkage estimator whose strength is governed P by within-panel disagreement D2 = N 1−1 i (δi − δ̄)2 : substituting the unbiased estimate τb2 = D2 into the oracle factor γ ⋆ = ∆2h /(∆2h +τ 2 /N ) yields a shrinkage weight that is monotone-decreasing in D2 , so D2 controls the ratio |δ ⋆ |/|∆| rather than |δ ⋆ | in isolation. A simpler but operationally important question, complementary to the proposition, is whether DISCA’s applied correction is in fact driven by disagreement at the scenario level, or whether it collapses into a per-country offset. Using Qwen2.5-7B as the backbone, we log per-scenario (D2 , |δ ⋆ |) pairs from the full DISCA pipeline across 5 countries (USA, JPN, DEU, VNM, ETH) and 310 MultiTP scenarios per country (n = 1,550 total), and compute Pearson’s correlation between log10 D2 and |δ ⋆ |. Result. The two quantities are positively and significantly correlated, r = +0.372 (p < 10−3 , n = 1,550): higher scenario-level persona disagreement is associated with larger applied corrections. This is consistent with the scenario-conditioned shrinkage policy formalised by Proposition 1, with the caveat that the proposition governs the ratio |δ ⋆ |/|∆| rather than the raw magnitude-a positive raw correlation arises when |∆| also co-varies with D2 across scenarios, which is the typical case when both are driven by the same underlying spread of persona logit gaps. At the country level (Table 16), the extremes match expectation-ETH, with the lowest mean variance (0.009), receives the smallest mean correction (0.009); USA, with the highest mean variance (0.086), receives the largest (0.015)-while the three middle countries (JPN, DEU, VNM) sit in a narrow band (0.012–0.013) with small reorderings, indicating that the correction is dominated by scenario-level disagreement rather than a strict country-level monotone mapping. Table 16: Per-country mean inter-persona variance D2 and mean correction magnitude |δ ⋆ | (n = 310 each). Countries are ordered by mean variance. The two extremes (ETH, USA) match expectation; the three middle countries sit in a narrow band, showing the correction is set primarily at the scenario level rather than by a strict country-level monotone mapping. Country
n
D2
|δ ⋆ |
ETH JPN DEU VNM USA
310 310 310 310 310
0.009 0.052 0.055 0.070 0.086
0.009 0.013 0.012 0.013 0.015
What this does and does not show. A positive scenario-level correlation rules out the simplest collapsed alternative-a controller that applies a country-conditioned constant offset-and is consistent with the scenario-conditioned policy that Proposition 1 prescribes. It is not a direct verification of the closed-form shrinkage in part (ii), which makes a quantitative claim about the ratio |δ ⋆ |/|∆| that a marginal scatter cannot identify without controlling for |∆|. We therefore present this diagnostic as evidence of scenario-level conditioning, not as a tight numerical test of the oracle shrinkage factor.
A10
Baseline Implementation Details
This appendix collects implementation details for every baseline reported in the paper. §A10.1 covers the training-free baselines of Table 1 (vanilla decoding, WVS Profile Prompt, PRISM-style framing, fixed logit offset). §A10.2 covers activation steering, and §A10.3 covers ARGS / controlled decoding. §A10.4 reports additional inference-time baselines run on the twenty-country Phi-4 grid (MC-Dropout, per-country temperature/margin scaling, and a DiffPO-style mixing baseline). 33
A10.1
Other training-free baselines (Table 1)
Vanilla decoding. The model is shown the native-language scenario with no persona prefix and no logit modification. We extract σ(δ(x)/Tdec ) and average across scenarios as in Eq. 7. WVS Profile Prompt. A single country-conditioned system prompt is constructed by concatenating the ten WVS descriptors of Table 26 (population aggregate, no cohort split) into one paragraph in the country’s native language; the model then answers each scenario under that prompt. This collapses the four-persona panel to a single point estimate, isolating the contribution of persona diversity on top of WVS grounding. PRISM-style prompt. Following the PRISM cultural-framing protocol of Kirk et al. [2024], we prepend a short cultural-context paragraph that names the country and asks the model to “answer as a typical respondent from [country] would”. No demographic stratification, no WVS data-a pure prompt-engineering control. Fixed logit offset. A country-specific scalar ∆c is added to the decision-token logits at inference (the same ∆c for every scenario in country c); ∆c is fit on the synthetic 200-scenario validation pool by gradient descent on JSD. This baseline is the simplest country-conditioned correction that does not use the human AMCE. A10.2
Activation Steering
For each country, a steering vector is computed as the difference in mean residual-stream activations from 20 culturally contrastive prompt pairs (e.g., “Answer as someone from [country] who values [WVS-high trait]” vs. “Answer as someone with the opposite values”). Vectors are extracted per backbone from the transformer midpoint of the same model the baseline is evaluated on (⌊L/2⌋ where L is the model’s layer count: layer 20 for Phi-4 (L=40), layer 40 for Llama-3.1-70B/3.370B (L=80), layer 20 for Magistral-Small-2509 (L=40)) and applied at inference time by adding α · vsteer to the residual stream, with α = 1.5 selected via the same synthetic validation set used for DISCA. There is no cross-model vector transfer; the dimension of vsteer matches the residual stream of the target backbone by construction. Sensitivity analysis and fairness considerations. Activation steering performance is known to be sensitive to (i) the extraction layer, (ii) the scaling coefficient α, and (iii) the construction of contrastive pairs [Arditi et al., 2025]. We conducted a limited sensitivity check on Llama-3.1-70B (the largest backbone, L=80): extracting from layers 32, 40, and 48 with α ∈ {0.5, 1.0, 1.5, 2.0, 3.0}, we found that layer 40 (the midpoint, used in Table 1) with α = 1.5 yielded the lowest mean JSD (0.0877); other configurations ranged from 0.089 to 0.112. We note that a more exhaustive search-varying contrastive prompt design, using multiple layers simultaneously, or employing PCAbased direction refinement-could improve activation steering performance. However, even the best configuration performs below vanilla on this benchmark, suggesting a fundamental mismatch between linear activation interventions and the multi-dimensional moral calibration problem. The key structural advantage of DISCA is that it operates per-scenario with adaptive correction magnitude, whereas activation steering applies a fixed direction uniformly across all scenarios. A10.3
ARGS and controlled decoding (relation to our work)
ARGS [Khanov et al., 2024] and controlled decoding [Mudgal et al., 2024] typically assume trained reward models; a cross-cultural deployment would require a separate reward per target country. On binary forced-choice MultiTP, one can instead define rewards as closed-form functionals of persona logit gaps; a Prospect-Theory kernel aligned with Eq. 3 is then the natural analogue of importanceweighted PT–IS. We do not run the full ARGS decoding stack in our experiments; Table 18 and §4.2 isolate the kernel comparison on the shared DISCA pipeline. A10.4
Additional inference-time baselines (twenty-country Phi-4 grid)
To isolate what within-country persona disagreement adds above simpler country-conditioned corrections, we compare DISCA against three inference-time alternatives on the same twenty-country Phi-4 grid. The temperature/margin scaling baselines are deliberately granted oracle access to each 34
country’s full human AMCE for hyperparameter selection (a strict super-set of any held-out calibration regime); MC-Dropout is country-agnostic. All numbers are computed on the same evaluation pool used in the main results, so the macro means line up exactly with Table 1 (Temp. 0.513, Margin 0.506) and Table 2 (DISCA 0.346). (6) MC-Dropout calibration. Following the moral uncertainty inflation template of Kwon et al. [2026], we enable every torch.nn.Dropout* module at inference (rate 0.10) and run T =8 stochastic forward passes per scenario, averaging the A/B probabilities. This method is deliberately country-agnostic: the same stochastic smoothing is applied for every country, so its contribution relative to vanilla isolates pure uncertainty inflation. (7) Per-country temperature / margin scaling (oracle). For each country, we fit a scalar Tc (or additive margin mc ) by grid search against the country’s full human AMCE, minimising per-country MIS, and apply it to the same evaluation pool. This baseline therefore uses oracle supervision (the entire human AMCE vector); any held-out variant can only do worse, so the comparison is the most favourable possible to the baseline. (8) DiffPO-binary. The spirit of DiffPO [Chen et al., 2025] adapted to binary decisions: we mix the vanilla probability with a country-conditioned target ptarget (c, cat) built directly from the country’s public human AMCE (paligned = (1 − α) pvan + α ptarget ), with α ∈ [0, 1] fit per country against the same AMCE. Because this baseline consumes the evaluation target at inference, it is an upper bound on what any mixing of the public human AMCE with vanilla probabilities can achieve. Takeaway. DISCA is compared against all three on the same twenty-country grid; because baselines (7) and (8) use oracle population-level supervision while DISCA uses only within-country persona disagreement (never touching the human AMCE), the comparison is deliberately favourable to the baselines. The per-country results are in Table 17: DISCA wins 18 of 20 countries against MC-Dropout, 19 of 20 against per-country temperature scaling, and 18 of 20 against margin scaling. Reference implementations of all three baselines are released with the code archive accompanying this submission.
Table 17: Phi-4 (14B) per-country head-to-head: DISCA vs. three inference-time baselines on the same 20-country test split. Per-country held-out MIS (↓, lower is better). Bold cells mark the best of {DISCA, MC-Dropout, Tc , mc } per country. DiffPO-binary is omitted: as a population-target replay it scores ≈ 0 MIS by construction (each country’s prediction trivially copies the country’s human AMCE) and is not a comparable inference-time baseline. Country
DISCA (ours)
MC-Dropout
Tc scale
Margin mc
ARG BGD BRA CHN COL DEU ETH GBR IDN IRN JPN KGZ MEX MMR MYS ROU SRB THA USA VNM
0.389 0.368 0.299 0.214 0.438 0.255 0.485 0.319 0.274 0.530 0.195 0.393 0.420 0.357 0.328 0.385 0.385 0.313 0.243 0.336
0.412 0.379 0.306 0.367 0.389 0.376 0.526 0.407 0.452 0.538 0.371 0.361 0.445 0.382 0.363 0.398 0.410 0.342 0.369 0.473
0.471 0.535 0.354 0.528 0.501 0.580 0.655 0.576 0.542 0.382 0.327 0.538 0.509 0.523 0.497 0.564 0.563 0.494 0.564 0.561
0.453 0.535 0.298 0.528 0.469 0.580 0.655 0.576 0.542 0.353 0.327 0.538 0.506 0.523 0.497 0.564 0.563 0.494 0.564 0.561
Macro Wins (DISCA vs.)
0.346 -
0.403 18/20
0.513 19/20
0.506 18/20
35
A11
Hyperparameters and Sensitivity Sweeps
This appendix collects the full hyperparameter specification of DISCA together with the sensitivity sweeps that justify the choices. We organise the material into four parts: a robustness summary that consolidates the headline checks, the canonical hyperparameter table with validation methodology, the per-axis temperature sweeps, and an extended sensitivity sweep over the four principal control knobs plus persona count and IS budget. A11.1
Robustness Summary
Table 18 consolidates the nine independent robustness checks referenced in §4.1. Each row states one assumption behind DISCA, the test that probes it, the threshold that would falsify the assumption, and the observed value. Per-axis sweeps and full per-country tables for each row appear in the subsections below and in App. A14. Table 18: Robustness suite. Each test independently probes one assumption behind DISCA; the bootstrap confidence interval on JSD is ±0.004. Acronyms: CS-Clamp = consensus-clamp deterministic shift (apply δ̄ directly, no IS); ARGS-Unif = ARGS-style controlled decoding [Khanov et al., 2024] with uniform reward over personas; ARGS-WVS = same with WVS-weighted reward; ARGS-PT = same with the Prospect-Theory kernel of Eq. 3.
A11.2
Claim
Test
Threshold Observed
Dataset preprocessing irrelevant Tdec insensitivity Uniform Tcat insensitivity Per-category Tcat insensitivity PT–IS > best deterministic shift WVS weighting alone insufficient PT–IS kernel ≡ ARGS-PT Cross-lingual pipeline generalises
JSD range over 6 configs JSD span (Sweep A) JSD span (Sweep B) JSD span (Sweep C) ∆JSD(CS-Clamp→PT–IS) ARGS-WVS ≻ ARGS-Unif ARGS-PT best/tied end-to-end runs
< 0.010 < 0.010 < 0.010 < 0.010 >0 < 8/15 ≥ 10/15 8/8
0.0034 0.0074 0.0073 0.0077 0.003 3/15 13/15 8/8
Default Hyperparameters and Validation
Table 19: DISCA hyperparameters. Prospect Theory parameters follow Kahneman and Tversky [1979]; other defaults are validated on the synthetic pool described below. Parameter
Symbol
Value
Persona agents IS samples / pass Reliability scale Perturbation std. dev. Cooperation weight IS softmax temp. ESS guard threshold PT curvature PT loss aversion Decision temp. Tcat (Species / Gender / other) Default logit divisor ESS anchor blend
N Khalf s σ λcoop η ρeff α=β κ Tdec Tlogit -
4 64 (total 2Khalf = 128) 0.04 0.3 0.7 0.5 0.1 0.88 2.25 0.5 4.0 / 3.5 / 1.5 3.0 ⋆ on (off: δ̄ + δIS only)
Implementation. The dual-pass controller follows the equations in §3. Optional environment overrides can change Khalf , reliability scale s, and whether the ESS-anchor blend is active; such flags must be applied before controller initialisation when running ablations. Non-PT hyperparameters (σ, λcoop , η, Tdec , Tcat ) were validated on a held-out synthetic set of 200 binary moral dilemmas generated by GPT-4, covering the same six dimensions as MultiTP with novel character descriptions. Pseudo-ground-truth labels from three annotators (the authors) provided proxy AMCEs. Grid search minimised mean JSD on Qwen2.5-72B. Transferability was verified 36
on a disjoint 50-scenario MultiTP English subset: performance within 1.2% relative JSD of oracletuned configuration. Tcat values reflect empirical logit magnitude distributions: Species and Gender produce systematically larger gaps, requiring higher temperatures. Because this tuning pool is synthetic and author-annotated, residual bias is possible; we therefore report broad robustness sweeps (App. A11, App. A14) to show conclusions are stable beyond one tuned point. A11.3
Temperature Sensitivity
We sweep the two temperature families on the (USA, DEU, JPN) subset of MultiTP using Llama-3.1-70B. Sweep A varies Tdec ∈ {0.10, 0.25, 0.50, 0.75, 1.0, 2.0}; Sweep B replaces the per-category temperatures by a single uniform Tcat ∈ {0.5, 1.0, 1.5, 2.0, 3.0, 4.0, 6.0}; Sweep C varies the “Others” bucket Tcat ∈ {0.75, 1.0, 1.5, 2.0, 3.0, 4.0} while holding Tcat [Species]=4.0, Tcat [Gender]=3.5. All three sweeps yield JSD spans below 0.010 (0.0074, 0.0073, 0.0077 respectively), and in every sweep the DISCA default (†) is strictly more conservative than the empirically best value (∆JSD ∈ {−0.0061, −0.0021, −0.0037}). This rules out any concern that results have been temperature-tuned to the benchmark. Table 20: Temperature sensitivity. Defaults (†) are strictly worse than the best setting in every sweep, providing a conservative lower bound. Sweep
Setting
JSD ↓
Pearson r ↑
∆JSD
A: Tdec
0.10 0.25 0.50† 0.75 1.00 2.00 (best)
0.0502 0.0491 0.0489 0.0476 0.0466 0.0428
0.662 0.658 0.630 0.636 0.641 0.627
+0.0013 +0.0002 −0.0013 −0.0023 −0.0061
0.0074
JSD span
B: Uniform Tcat
0.5 1.0 1.5† 2.0 3.0 4.0 6.0 (best)
0.0521 0.0492 0.0469 0.0461 0.0454 0.0451 0.0448
+0.0052 +0.0023 −0.0008 −0.0015 −0.0018 −0.0021
0.0073
JSD span
C: Others Tcat
0.642 0.673 0.700 0.707 0.707 0.701 0.684
0.75 1.00 1.50† 2.00 3.00 4.00 (best)
0.0526 0.0512 0.0487 0.0475 0.0458 0.0450
0.566 0.589 0.634 0.660 0.693 0.706
+0.0039 +0.0025 −0.0012 −0.0029 −0.0037
0.0077
JSD span
The monotonic JSD reduction with larger Tdec reflects the role of Tdec in undoing RLHF logit compression: higher values widen the effective token-probability landscape so that IS perturbations exert larger directional influence. However, r peaks near Tdec =1.0 and declines at 2.0, indicating a tradeoff between distributional alignment (JSD) and rank-order alignment (r); we keep Tdec =0.5 as a deliberately conservative operating point. A11.4
Extended Hyperparameter Sensitivity
We probe the sensitivity of DISCA to the four principal hyperparameters: s (reliability gate scale), λcoop (individual-vs-consensus weight), σ (IS proposal floor), and Tcat (uniform logit-temperature scaling). For each axis we sweep five values on a three-country panel (USA, VNM, DEU) and hold the other three axes at their defaults; the s axis directly probes how the choice of reliability scale in Eq. 5 affects downstream MIS. 37
Table 21: Hyperparameter sensitivity ranges on the three-country panel (Phi-4, n=250/country). Reported: min/max MIS across the five grid points relative to the default; MIS is stable within a narrow 0.0037–0.0090 span on every axis, so the gains in Table 2 are not the product of a particular hyperparameter choice. Axis
Default
Grid
s λcoop σ Tcat scale
0.04 0.70 0.30 1.0×
{0.01, 0.02, 0.04, 0.08, 0.16} {0.30, 0.50, 0.70, 0.85, 0.95} {0.10, 0.20, 0.30, 0.45, 0.60} {0.33×, 0.67×, 1.0×, 1.33×, 1.67×}
Min MIS Max MIS 0.4091 0.4085 0.4084 0.4060
0.4164 0.4173 0.4122 0.4149
∆ 0.0073 0.0088 0.0037 0.0090
Why these four axes. s controls the steepness of the reliability gate (Eq. 5); λcoop trades off individual-persona PT utility against the consensus utility; σ floors the IS proposal scale so the N =4 empirical std cannot collapse to zero; and Tcat scales the whole family of per-category decision temperatures (preserving relative balance)-a global sharpening / flattening knob on the decision logits. Per-category temperatures are stress-tested separately in the temperature sensitivity sweep above.
Persona count, IS budget, and per-category vs. global Tcat . We additionally swept three operational knobs on the same USA/VNM/DEU panel. Persona count (N ∈ {2, 3, 4, 5, 6}, Table 22): N =4 is optimal at macro MIS 0.4051, with degradation at both smaller and larger panels (0.4252 at N =2, 0.4179 at N =6), supporting the four-persona design. IS budget (Khalf ∈ {8, 16, 32, 64, 128, 192}, Table 23): performance is flat across an order of magnitude (best 0.4092 at K=16; 0.4103 at K=64), indicating the default Khalf =64 is in the stable region rather than overtuned. Global vs. per-category Tcat (Table 24): the per-category default outperforms every global schedule we tested (0.4098 vs. all seven global values in {1.0, 1.5, 2.0, 2.5, 3.0, 3.5, 4.0}, which land in 0.4119–0.4138). The advantage is small in absolute terms but consistent.
Table 22: Persona-count sweep on the USA/VNM/DEU panel (Phi-4, n=250/country). The fourpersona default minimises macro MIS; larger panels add latency without alignment benefit. N
Macro MIS ↓
Std (across countries)
Flip rate
sec/scenario
2 3 4 5 6
0.4252 0.4194 0.4051 0.4107 0.4179
0.1045 0.0945 0.0829 0.0745 0.0919
0.275 0.273 0.276 0.276 0.245
0.089 0.100 0.111 0.120 0.130
Table 23: IS-budget sweep (Khalf samples per pass; total per scenario is Ktotal =2Khalf ). Macro MIS is flat across an order of magnitude. The mean reliability weight r̄ (gate column) grows monotonically with K, confirming the gate becomes more confident as the IS proposal cloud densifies. Khalf
Ktotal
Macro MIS ↓
Flip rate
r (gate)
sec/scen
8 16 32 64 128 192
16 32 64 128 256 384
0.4114 0.4092 0.4104 0.4103 0.4147 0.4126
0.284 0.271 0.278 0.280 0.274 0.273
0.684 0.736 0.766 0.793 0.814 0.827
0.121 0.114 0.114 0.113 0.116 0.112
38
Table 24: Global vs. per-category Tcat on the USA/VNM/DEU panel. The per-category default (decision 4.0, social 3.5, non-social 1.5) outperforms every global value we tested. The gap is small (≤ 0.004) but uniformly directional, supporting category-specific decision temperatures over a single shared scalar. Macro MIS ↓
Std
Flip rate
per-category default (4.0 / 3.5 / 1.5)
0.4098
0.0881
0.278
global Tcat = 1.00 global Tcat = 1.50 global Tcat = 2.00 global Tcat = 2.50 global Tcat = 3.00 global Tcat = 3.50 global Tcat = 4.00
0.4120 0.4119 0.4138 0.4134 0.4127 0.4138 0.4128
0.0921 0.0838 0.0778 0.0774 0.0749 0.0689 0.0734
0.292 0.304 0.301 0.296 0.303 0.306 0.302
Tcat schedule
Per-persona utility floor (minority safeguard). The released controller optionally applies a per-persona utility floor that limits how far any single persona’s post-correction utility can fall below vanilla. Sweeping f ∈ {0.0, 0.5, 1.0, 2.0} on the same USA/VNM/DEU panel (Phi-4, n=250/country) keeps macro MIS in [0.4046, 0.4157]—well inside the macro-MIS bootstrap noise floor-so the floor is a stability knob rather than a load-bearing hyperparameter. Table 25: Per-persona utility floor sweep (USA / VNM / DEU). Macro MIS is flat across two orders of magnitude of the floor parameter.
A12
Floor f
Macro MIS ↓
Std (across countries)
0.0 (off) 0.5 1.0 2.0
0.4083 0.4046 0.4066 0.4157
0.0717 0.0696 0.0751 0.0684
Persona Construction Details
This appendix specifies how the four cultural personas per country are constructed and provides empirical support that the construction is grounded in human survey data rather than authored ad hoc. §A12.1 reproduces two personas verbatim from generate_wvs_persona(); §A12.2 documents the WVS Wave 7 processing pipeline; §A12.3 links the ten WVS cultural dimensions to the six MultiTP moral attributes both thematically and empirically; and §A12.5 shows that DISCA is robust to the choice of the fourth (anchor) persona. A12.1
Persona Prompts
Each country yields four personas: three age-cohort personas (young, middle, older) generated from the country-and-cohort-specific WVS-7 means, plus a population-wide aggregate that serves as a demographic anchor. Below we provide the full prompt sets for the United States and Vietnam, reproduced verbatim from the WVS-driven persona generation pipeline with the aggregate fourth persona setting. United States (all four personas, verbatim) Young cohort. You are a young adult from the United States, in your 20s and early 30s. Your worldview is shaped by the cultural values prevalent in your community. On matters of faith you are moderately religious. On raising children you are firmly oriented toward independence and imagination. On contested moral choices you are morally conservative on contested issues. In your dealings with strangers you have a guarded attitude toward strangers. Civically you are an active political participant who signs petitions, joins boycotts and takes part in lawful demonstrations. You are moderately proud of your country. Overall you are rather happy with your life. On the role of women in society you are moderately egalitarian
39
on gender roles. In what you prioritise in life you are leaning materialist, prioritising economic and physical security. Toward people unlike yourself you are highly tolerant of outgroups such as immigrants, minorities and people with different lifestyles. When you face a moral dilemma, you weigh the choices through this set of values and answer in a way that is consistent with the worldview above.
Middle-aged cohort. You are a middle-aged adult from the United States, in your 40s or 50s. Your worldview is shaped by the cultural values prevalent in your community. On matters of faith you are moderately religious. On raising children you are firmly oriented toward independence and imagination. On contested moral choices you are strictly opposed to such contested moral acts. In your dealings with strangers you have a guarded attitude toward strangers. Civically you are an active political participant who signs petitions, joins boycotts and takes part in lawful demonstrations. You are moderately proud of your country. Overall you are rather happy with your life. On the role of women in society you are moderately egalitarian on gender roles. In what you prioritise in life you are leaning materialist, prioritising economic and physical security. Toward people unlike yourself you are highly tolerant of outgroups such as immigrants, minorities and people with different lifestyles. When you face a moral dilemma, you weigh the choices through this set of values and answer in a way that is consistent with the worldview above.
Older cohort. You are a senior citizen from the United States, over 60 years old. Your worldview is shaped by the cultural values prevalent in your community. On matters of faith you are moderately religious. On raising children you are firmly oriented toward independence and imagination. On contested moral choices you are strictly opposed to such contested moral acts. In your dealings with strangers you have a guarded attitude toward strangers. Civically you are an active political participant who signs petitions, joins boycotts and takes part in lawful demonstrations. You are intensely proud of your country. Overall you are rather happy with your life. On the role of women in society you are moderately egalitarian on gender roles. In what you prioritise in life you are leaning materialist, prioritising economic and physical security. Toward people unlike yourself you are highly tolerant of outgroups such as immigrants, minorities and people with different lifestyles. When you face a moral dilemma, you weigh the choices through this set of values and answer in a way that is consistent with the worldview above.
Aggregate cohort (all ages). You are a adult citizen from the United States. Your worldview is shaped by the cultural values prevalent in your community. On matters of faith you are moderately religious. On raising children you are firmly oriented toward independence and imagination. On contested moral choices you are strictly opposed to such contested moral acts. In your dealings with strangers you have a guarded attitude toward strangers. Civically you are an active political participant who signs petitions, joins boycotts and takes part in lawful demonstrations. You are moderately proud of your country. Overall you are rather happy with your life. On the role of women in society you are moderately egalitarian on gender roles. In what you prioritise in life you are leaning materialist, prioritising economic and physical security. Toward people unlike yourself you are highly tolerant of outgroups such as immigrants, minorities and people with different lifestyles. When you face a moral dilemma, you weigh the choices through this set of values and answer in a way that is consistent with the worldview above.
Vietnam (all four personas, Vietnamese, verbatim) Young cohort. Bạn là một thanh niên đến từ Việt Nam, ở độ tuổi 20 đến đầu 30. Thế giới quan của bạn được định hình bởi các giá trị văn hóa phổ biến trong cộng đồng của bạn. Về vấn đề đức tin, bạn khá thế tục. Trong việc nuôi dạy con cái, bạn nghiêng về sự vâng lời và đức tin tôn giáo. Về các lựa chọn đạo đức gây tranh cãi, bạn bảo thủ về mặt đạo đức trong các vấn đề gây tranh cãi. Trong giao tiếp với người lạ, bạn có sự ngờ vực sâu sắc đối với người khác. Về mặt công dân, bạn là một người tham gia chính trị thụ động. Bạn rất tự hào về đất nước của bạn. Nhìn chung, bạn rất hài lòng với cuộc sống của bạn. Về vai trò của phụ nữ trong xã hội, bạn khá truyền thống về vai trò giới. Về những gì bạn ưu tiên trong cuộc sống, bạn
40
có xu hướng hậu vật chất. Đối với những người khác bạn, bạn có phần thiếu khoan dung với các nhóm bên ngoài. Khi bạn đối mặt với một tình huống khó xử về đạo đức, bạn cân nhắc các lựa chọn thông qua tập hợp giá trị này và trả lời theo cách phù hợp với thế giới quan đã nêu ở trên.
Middle-aged cohort. Bạn là một người trung niên đến từ Việt Nam, ở độ tuổi 40 hoặc 50. Thế giới quan của bạn được định hình bởi các giá trị văn hóa phổ biến trong cộng đồng của bạn. Về vấn đề đức tin, bạn khá thế tục. Trong việc nuôi dạy con cái, bạn nghiêng về sự vâng lời và đức tin tôn giáo. Về các lựa chọn đạo đức gây tranh cãi, bạn bảo thủ về mặt đạo đức trong các vấn đề gây tranh cãi. Trong giao tiếp với người lạ, bạn có thái độ dè dặt với người lạ. Về mặt công dân, bạn là một người thỉnh thoảng tham gia chính trị. Bạn rất tự hào về đất nước của bạn. Nhìn chung, bạn rất hài lòng với cuộc sống của bạn. Về vai trò của phụ nữ trong xã hội, bạn khá truyền thống về vai trò giới. Về những gì bạn ưu tiên trong cuộc sống, bạn có xu hướng hậu vật chất. Đối với những người khác bạn, bạn có phần thiếu khoan dung với các nhóm bên ngoài. Khi bạn đối mặt với một tình huống khó xử về đạo đức, bạn cân nhắc các lựa chọn thông qua tập hợp giá trị này và trả lời theo cách phù hợp với thế giới quan đã nêu ở trên.
Older cohort. Bạn là một người cao tuổi đến từ Việt Nam, trên 60 tuổi. Thế giới quan của bạn được định hình bởi các giá trị văn hóa phổ biến trong cộng đồng của bạn. Về vấn đề đức tin, bạn có niềm tin tôn giáo ở mức trung bình. Trong việc nuôi dạy con cái, bạn nghiêng về sự vâng lời và đức tin tôn giáo. Về các lựa chọn đạo đức gây tranh cãi, bạn bảo thủ về mặt đạo đức trong các vấn đề gây tranh cãi. Trong giao tiếp với người lạ, bạn có thái độ dè dặt với người lạ. Về mặt công dân, bạn là một người thỉnh thoảng tham gia chính trị. Bạn rất tự hào về đất nước của bạn. Nhìn chung, bạn rất hài lòng với cuộc sống của bạn. Về vai trò của phụ nữ trong xã hội, bạn khá truyền thống về vai trò giới. Về những gì bạn ưu tiên trong cuộc sống, bạn có xu hướng hậu vật chất. Đối với những người khác bạn, bạn rất thiếu khoan dung với các nhóm bên ngoài. Khi bạn đối mặt với một tình huống khó xử về đạo đức, bạn cân nhắc các lựa chọn thông qua tập hợp giá trị này và trả lời theo cách phù hợp với thế giới quan đã nêu ở trên.
Aggregate cohort (all ages). Bạn là một công dân trưởng thành đến từ Việt Nam. Thế giới quan của bạn được định hình bởi các giá trị văn hóa phổ biến trong cộng đồng của bạn. Về vấn đề đức tin, bạn khá thế tục. Trong việc nuôi dạy con cái, bạn nghiêng về sự vâng lời và đức tin tôn giáo. Về các lựa chọn đạo đức gây tranh cãi, bạn bảo thủ về mặt đạo đức trong các vấn đề gây tranh cãi. Trong giao tiếp với người lạ, bạn có thái độ dè dặt với người lạ. Về mặt công dân, bạn là một người thỉnh thoảng tham gia chính trị. Bạn rất tự hào về đất nước của bạn. Nhìn chung, bạn rất hài lòng với cuộc sống của bạn. Về vai trò của phụ nữ trong xã hội, bạn khá truyền thống về vai trò giới. Về những gì bạn ưu tiên trong cuộc sống, bạn có xu hướng hậu vật chất. Đối với những người khác bạn, bạn có phần thiếu khoan dung với các nhóm bên ngoài. Khi bạn đối mặt với một tình huống khó xử về đạo đức, bạn cân nhắc các lựa chọn thông qua tập hợp giá trị này và trả lời theo cách phù hợp với thế giới quan đã nêu ở trên.
Aggregate (fourth) persona. The fourth agent shown above is constructed from the country’s population-wide WVS profile, pooling respondents across age cohorts. It serves as a demographic anchor: by suppressing cohort-specific noise it ensures that the ensemble’s consensus target δ̄ reflects the broadest available empirical signal rather than a single generational viewpoint. §A12.5 contrasts this choice with a country-invariant utilitarian anchor and shows that DISCA’s macro-MIS is insensitive to the substitution. Faithfulness disclaimer. The descriptors above are read directly from the country’s WVS-7 means through deterministic quartile cuts (Table 26); they are not hand-selected to match cultural stereotypes. As a consequence, individual cohorts can present empirically grounded but locally counterintuitive combinations (for example, a U.S. young-adult that is leaning materialist and morally conservative on contested issues), reflecting actual WVS-7 distributions rather than authorial bias. 41
A12.2
WVS Data Processing Pipeline
We process WVS Wave 7 individual-level microdata following the ten-variable cultural-value scheme of Greco et al. [2026]. The pipeline is fully deterministic: given the inverted WVS-7 CSV and the country code, it returns the four personas used at inference time without any tunable hyperparameters. 1. Variable extraction: For each respondent, we extract 10 value dimensions from WVS items (Table 26). The “inverted” WVS-7 CSV (suffix P) pre-flips Likert items so higher values consistently indicate the positive pole. Items without the P suffix (Q152, Q153, Q177–Q182) use original WVS coding. 2. Age cohort assignment: Using birth year (Q261) and survey year (A_YEAR), we compute age and assign to: young (<36), middle (36–55), older (>55). Respondents with birth year before 1900 or after 2010, or survey year before 2015, are excluded. Negative WVS codes (−1, −2, −4, −5: refusal / don’t know) are dropped; zero is retained for binary 0/1 items. 3. Cohort aggregation: For each country and age cohort, each dimension’s score is the unweighted mean of all valid (respondent, item) values associated with that dimension - i.e., for multi-item dimensions (e.g., Q177–Q182 for moral acceptability) every respondent contributes one value per item to a single pooled mean. For single-item dimensions this reduces to the standard across-respondent mean. 4. Normalisation and descriptor mapping: Raw dimension means are normalised to [0, 1] via (value − lo)/(hi − lo) where (lo, hi) is the WVS scale range, then directionally flipped so higher = positive pole. Four-level descriptors are assigned by quartile cuts: ≥ 0.75 (strong positive), ≥ 0.50, ≥ 0.25, < 0.25 (strong negative). See Table 26. 5. Fallback: For countries with insufficient WVS coverage (e.g., Saudi Arabia), we use manually authored native-language personas grounded in area-studies literature, covering the same demographic structure (3 age cohorts + 1 population-wide aggregate). Table 26: The 10 WVS cultural-value dimensions used for persona generation, following Greco et al. [2026]. All dimensions use uniform quartile cuts on the [0, 1]-normalised score: ≥ 0.75 (descriptor 1, strongest positive pole), ≥ 0.50 (2), ≥ 0.25 (3), < 0.25 (4, strongest negative pole). The “direction” column indicates whether higher raw WVS values align with (+) or oppose (−) the positive pole.
A12.3
Dimension
WVS items
Range Dir. Descriptors (≥0.75 / ≥0.50 / ≥0.25 / <0.25)
Religiosity Child-rearing Moral accept. Social trust Polit. particip. National pride Happiness Gender equality Materialism Tolerance
Q6P Q17P Q177–Q182 Q57P Q199P, Q200P Q254P Q46P Q29P, Q30P, Q31P, Q33P Q152, Q153 Q19P–Q23P
1–4 0–1 1–10 1–2 1–3 1–4 1–4 1–4 1–3 0–1
+ − + + + + + − + −
deeply / moderately religious / somewhat / highly secular independence-imagination ↔ obedience-faith very permissive ↔ strictly opposed very high trust ↔ deep distrust active participant ↔ non-participant very proud ↔ not proud very happy ↔ not happy strongly egalitarian ↔ traditional roles post-materialist ↔ materialist very tolerant ↔ intolerant
WVS-to-Trolley Dimension Linkage
The linkage between the ten WVS cultural dimensions and the six MultiTP moral dimensions operates through two complementary mechanisms. Direct thematic overlap. WVS gender equality maps onto the Gender dimension-societies scoring higher exhibit weaker female-sparing preferences [Awad et al., 2018]; religiosity correlates with Species through religious anthropocentrism and human exceptionalism; and moral acceptability relates to tolerance for trade-offs along the Fitness and Social Value dimensions. Indirect modulation. Dimensions such as social trust, child-rearing values, and materialism orientation reshape the moral frame the persona adopts: post-materialist personas favour utilitarian calculi, while obedience-oriented child-rearing values up-weight hierarchical social roles (affecting Age and Social Value). Importantly, DISCA does not require an exact causal mapping between 42
WVS and MultiTP dimensions; the ten-dimensional persona ensemble produces diversity in logitspace gaps, and the importance-sampling stage selects the correction that balances collective utility. Empirical validation of the WVS→AMCE pathway. Table 27 regresses each of the six MultiTP human AMCE dimensions on the 10 WVS cultural features across the 20-country panel (standardised OLS). The 10-feature WVS profile explains R2 ∈ [0.55, 0.69] of human AMCE variance, with |β| > 0.3 in 31 of 60 cells. This is direct evidence that WVS-grounded persona construction is not arbitrary: the cultural axes our personas inherit correlate strongly with the moral preferences DISCA must reproduce. Table 27: WVS features as predictors of human AMCE dimensions. OLS with standardised pre2 dictors. Coefficients are standardised (β); bold = |β| > 0.3. Radj measures how well the 10 WVS cultural dimensions explain each moral preference dimension across the 20-country panel. AMCE dim
relig
natio happi gende mater
toler
2 R2 Radj
Species Gender Age Fitness SocialValue Utilitarianism
+0.47 +0.12 -0.65 +0.60 +0.08 +0.16 -0.02 -0.74 -0.30 +0.69 +0.37 -0.18 +0.41 -0.44 -0.39 +0.71 -0.81 -0.20 -0.80 -0.09 +0.61 -0.74 +0.17 +0.26 -0.27 +0.62 +0.49 +0.28 -0.23 -0.16 -0.12 -0.17 +0.43 +0.09 -1.46 -0.21 -0.27 -0.30 +0.40 -0.40 -0.37 +0.58 +0.17 -0.31 -0.12 +1.11 +0.39 +0.07 +0.59 -0.21 -0.01 +0.13 -0.76 +0.42
+0.40 +0.06 -0.37 +0.53 -0.15 +0.25
0.57 0.66 0.55 0.60 0.69 0.66
child moral
socia
polit
0.09 0.29 0.05 0.15 0.35 0.29
Table 28 closes the loop with a causal leave-one-WVS-dim-out probe: for each WVS dimension d, we re-run DISCA with d removed from every persona and measure the change in per-MultiTPdim absolute error (positive cell = error rises when the dimension is removed = that dimension is load-bearing for the corresponding moral attribute). The three largest positive couplings are moral acceptability → Social Value (+3.97 pp), social trust → Social Value (+2.81 pp), and child-rearing values → Utilitarianism (+2.41 pp). Cells with the top-3 largest positive couplings per row are bolded; rows with fewer than three positive cells are bolded only on those positive cells. Negative cells are also informative: tolerance of diversity → Social Value (−7.76) and gender equality → Gender (−3.42) indicate that, for those moral attributes, the corresponding WVS descriptor biases the persona ensemble away from the human target, so removing the descriptor improves accuracy. Crucially, DISCA does not require manual feature selection to handle these noisy dimensions: the loss-averse importance-sampling stage (Eq. 3) automatically down-weights candidate perturbations that worsen any persona’s alignment, effectively acting as a built-in noise filter. When a WVS dimension introduces bias for a particular moral attribute, the PT-IS utility penalises the resulting correction asymmetrically (κ=2.25× for losses vs. gains), ensuring that noisy persona signals are suppressed rather than propagated. This self-correcting property is why the method remains robust despite imperfect WVS-to-trolley mappings-a design-level answer to the concern that the linkage is “indirect modulation” rather than causal. Table 28: WVS-dim × MultiTP-dim causal impact matrix. Cells are the mean increase in per-dim AMCE error (pp) when the WVS dimension is dropped from every persona, macro-averaged across the country panel. Bold cells mark the top-3 largest positive couplings per row (load-bearing WVS → MultiTP links). WVS dim religiosity child rearing moral acceptability social trust political participation national pride happiness gender equality materialism orientation tolerance diversity
A12.4
Species_Humans
Gender_Female
Age_Young
Fitness_Fit
SocialValue_High
Utilitarianism_More
+0.66 -1.82 +1.13 +0.11 -1.24 -0.37 -0.23 +0.41 +1.35 -0.76
+1.65 +1.29 -0.02 +1.43 -0.82 -0.10 -0.88 -3.42 -1.26 -3.52
-1.66 +0.86 -1.56 -0.85 -0.63 -0.28 +0.13 -2.90 +0.90 +1.95
+0.58 +1.30 -0.34 +0.51 -0.36 +0.52 -1.69 -1.11 +0.04 -0.59
+1.02 -0.33 +3.97 +2.81 -0.16 +1.19 -0.11 -1.87 -0.36 -7.76
-0.37 +2.41 -1.69 -1.88 +0.86 +1.21 +1.72 +0.62 +2.15 +1.54
Macro-level WVS dimension importance
The per-MultiTP-dim impact matrix above (Table 28) tells us which moral attribute each WVS descriptor influences. A coarser but operationally useful question is which WVS dimensions are 43
load-bearing for macro alignment, taken as a whole. We answer this with a leave-one-WVS-dimout probe: for each of the ten descriptors, we re-run DISCA with that descriptor removed from every persona and record the change in macro MIS, averaged across USA, JPN, and VNM as a representative panel. The result is a concentrated importance profile: religiosity, gender equality, and moral acceptability each raise MIS by more than 0.03 when removed; national pride and happiness barely register (< 0.01). The implication for persona engineering is that the ten-dimensional WVS profile can be compressed substantially while preserving most of the steering signal, which is consistent with the IS stage’s built-in down-weighting of noisy descriptors discussed in App. A12.3. Table 29: WVS dimension dropout (3-country average; higher ∆MIS means more load-bearing). Dimension names follow the canonical WVS-7 list of Table 26 and Table 28.
A12.5
Dropped WVS dimension
∆MIS (avg)
religiosity gender equality moral acceptability social trust child rearing political participation materialism orientation tolerance diversity national pride happiness
+0.046 +0.041 +0.034 +0.022 +0.019 +0.017 +0.013 +0.011 +0.007 +0.005
Sensitivity to the Fourth Persona
Setup. App. A12 pairs three age-cohort personas with a population-wide aggregate WVS profile as the fourth agent. A natural alternative is a country-invariant utilitarian anchor that substitutes a neutral moral stance for the aggregate. To confirm that this choice is not load-bearing, we ran a head-to-head comparison of the two variants on the twenty-country Phi-4 grid. Result. The utilitarian-anchor variant improves macro MIS from 0.4252 to 0.4049 (∆ = −0.0203), a marginal gain that does not change any qualitative conclusion of the paper. Decision. We retain the aggregate variant as the default to preserve strict country-grounding of every persona and to avoid importing a globally fixed moral prior into country-conditioned steering. The utilitarian variant remains available as an opt-in for practitioners who prefer a neutral anchor (build_country_personas(..., fourth=‘utilitarian’)).
A13
AMCE Estimation Details
Uniform treatment of all dimensions. All six dimensions-Species, Gender, Age, Fitness, Social Value, and Utilitarianism-are computed identically. For each dimension, we compute the empirical mean of pspare (x) across all scenarios in that dimension (Eq. 7). For the five binary dimensions, each scenario presents a forced choice between two character groups differing on exactly one attribute. For Utilitarianism, scenarios vary the number of lives on each side; the binary indicator is whether the larger group (“More”) or the smaller group (“Less”) is spared. This uniform mean-based estimation is consistent with the compute_ACME method in the MultiTP codebase [Jin et al., 2025], which fits a no-intercept linear regression with a binary group indicator-mathematically equivalent to taking the mean of the saving probability for the preferred group. Human AMCE scale. The MultiTP ground-truth AMCEs in the MultiTP-released countryspecific AMCE table is on a [−1, 1] scale. We convert to [0, 100] via (1 + AMCEraw )/2 × 100, where 50% represents no preference and 100% represents maximal preference for the “preferred” group. Model AMCEs are computed on the same scale. Two reporting scales (resolving an apparent contradiction). The PCA visualisation in Figure 3 reports ℓ2 distances on the [0, 100] percentage-point scale (e.g., 83.3 → 64.6), while every numerical 44
table in the main text and appendix (Tables 1, 2, 4, 7, etc.) reports MIS on the [0, 1] proportional scale (i.e., directly in pspare space, with values in [0.2, 1.0]). The two scales differ by a factor of 100. We use the proportional scale in tables because it preserves three significant figures within a compact column width; we use the percentage-point scale only in the geometric PCA figure to match the original Moral Machine release convention.
Consistency. Awad et al. [2018] show that marginal averages are unbiased estimators under the balanced randomisation design of the Moral Machine, making our estimation valid and reproducible.
Cross-lingual A/B token elicitation. Eq. 7 requires a clean A vs. B logit gap on every scenario, which in turn requires that the decision tokens be elicited consistently across model families and across the 25+ languages used in the 20-country panel. Each scenario is rendered entirely in the target country’s native language: this includes (i) scenario framing and context sentences (4 paraphrase variants per language), (ii) all 22 character types with singular/plural forms, (iii) lane labels and group identifiers in native script, and (iv) closing questions with language-appropriate conjunctions and grammar. The prompt frame wraps the scenario in a native-language instruction that explicitly requests an English answer token, following the pattern: “[moral dilemma preamble in native language] \n {scenario} \n [instruction to answer A or B in English]”. This keeps moral reasoning in the culturally native linguistic frame while maintaining a single decision interface across the panel.
Token verification. Token IDs are verified per model at initialisation by encoding the uppercase strings "A" and "B" with add_special_tokens=False and asserting single-token mappings: Qwen2.5: A=22397, B=11572; Llama-3.1: A=31266, B=9268. Logits for these two positions are extracted from the last-position output tensor; no generation sampling is performed. The model uses its built-in chat_template for correct role formatting across model families.
A14
Dataset Preprocessing and Sensitivity
Pipeline. Each MultiTP CSV is processed in five steps: (i) deduplication keeps only the canonical paraphrase (which_paraphrase = 0); (ii) the Utilitarianism quality filter drops scenarios where the two groups have identical size and all characters belong to the quality-attribute role set {P REGNANT, W OMAN, L ARGE W OMAN}, which would otherwise inject a non-numerosity signal into a dimension whose AMCE is by construction the count-difference effect; (iii) per-category capping retains at most 80 scenarios per dimension; (iv) under-represented categories are oversampled with replacement to a minimum of 36 scenarios; and (v) the resulting pool is shuffled with a fixed seed (42).
Reproducible left/right assignment. When the MultiTP paraphrase_choice field does not unambiguously determine which group is rendered on the left vs. right, we deterministically fall back to a SHA-256 hash of the tuple (sub1 , sub2 , g1 , g2 ) modulo 2, so the same scenario receives the same ordering across runs and machines. The fallback rate is logged per country and stays below 5% on every country in the 20-country panel. Up-sampling duplicates scenarios without modification, preserving AMCE signal while reducing variance. Category capping prevents dimension dominance. Side balancing enables effective debiasing. To evaluate whether these choices alter the effective evaluation distribution, we compare six preprocessing configurations on 10 representative countries (USA, DEU, CHN, JPN, BRA, VNM, GBR, KOR, RUS, NGA) using Llama-3.1-70B. 45
Table 30: Per-country dataset sensitivity. JSD range = 0.0034 (aggregate) and < 0.010 in 9/10 countries; the only exception is RUS at 0.013, driven by augmentation sensitivity in the rare Species/Utilitarianism categories. D0 is retained as default. Country
D0-Default
D1-NoAug
D2-Cap40
D3-Cap120
D4-NoFlip
D5-Strict
USA DEU CHN JPN BRA VNM GBR KOR RUS NGA
0.0655 0.0465 0.0393 0.0304 0.0492 0.0592 0.0614 0.0371 0.0383 0.0522
0.0674 0.0515 0.0363 0.0291 0.0551 0.0574 0.0667 0.0354 0.0489 0.0568
0.0625 0.0492 0.0385 0.0319 0.0479 0.0631 0.0608 0.0378 0.0389 0.0506
0.0644 0.0467 0.0383 0.0296 0.0510 0.0585 0.0618 0.0357 0.0374 0.0525
0.0643 0.0444 0.0409 0.0273 0.0520 0.0578 0.0597 0.0349 0.0383 0.0514
0.0684 0.0484 0.0398 0.0320 0.0480 0.0626 0.0633 0.0382 0.0357 0.0538
Mean ∆ vs D0
0.0479 -
0.0505 +0.0025
0.0481 +0.0002
0.0476 −0.0003
0.0471 −0.0008
0.0490 +0.0011
D1-NoAug shows the largest degradation (+0.0025), justifying synthetic augmentation of the underrepresented Species and Utilitarianism categories. D4-NoFlip is marginally lower than D0, indicating that group-side randomisation is conservative (it adds noise) rather than harmful; we retain it because it preserves positional-debiasing integrity. The 0.0034 aggregate range is well below the ±0.004 bootstrap CI, so reported results are not confounded by preprocessing choices.
A15
Trigger Mechanism and Latency
DISCA runs dual-pass IS on every scenario; inter-persona variance is logged but does not gate sampling. The flip rate (IS reversals vs. consensus) is 29.0% in the cross-backbone ablation suite (Table 4), indicating confidence modulation rather than wholesale decision reversal. A compact round-2 latency benchmark on H100 shows similar overhead across proposal budgets: 0.086s/scenario at K=64, 0.083s at K=128, and 0.083s at K=256, compared with 0.023s for vanilla decoding (3.57–3.72× overhead).
Table 31: Inference latency benchmark (Phi-4, H100). Method
K
sec/scenario
Overhead
DISCA DISCA DISCA vanilla
64 128 256 –
0.086 0.083 0.083 0.023
3.72× 3.59× 3.57× 1.00×
Cost-vs-quality frontier across backbones. Figure 5 plots the deployed-cost side of the Phi-4vs-Llama narrative: at 350 ms per scenario, Phi-4 with DISCA reaches MIS = 0.346, while Llama3.3-70B (with DISCA) takes 1414 ms-roughly 4× slower-and still ends at MIS = 0.668. Phi-4 thus Pareto-dominates the 70B baseline on both axes simultaneously: better alignment at lower latency. Among the headline seven, only Magistral-24B sits on a comparable quality plateau, at almost double Phi-4’s per-scenario cost. 46
Cost-vs-quality frontier (marker size ∝ params, color ∝ ΔMIS) Llama-3.3-70B (70.6B, Δ=+0.182)
DISCA mean MIS ( ↓ better)
0.65
0.60
Phi-3.5-Mini (3.8B, Δ=+0.067)
0.55
0.50 Qwen3-VL-8B (8.0B, Δ=+0.109)
0.45
0.40
0.35
Qwen2.5-7B (7.6B, Δ=+0.091)
Magistral-24B Phi-4 (14B) (14.7B, Δ=+0.108)(24.0B, Δ=+0.053)
103
DISCA per-scenario latency (ms, log)
Figure 5: Cost-vs-quality frontier on the headline 7 models. Per-scenario DISCA latency (log scale) vs. mean DISCA MIS. Marker size is proportional to parameter count; color encodes ∆MIS (greener = larger DISCA gain). Phi-4 (14B, ∆ = +0.108) lies bottom-left: Pareto-dominant over Llama-3.3-70B in both latency and alignment.
A16
Relationship to Persona-Dependent LLM Alignment
Kim et al. [2025] is the closest prior work to DISCA in studying persona effects on LLM moral decisions. Their study and ours share the Moral Machine framework and AMCE-based evaluation, but differ fundamentally in goal, method, and scope. Table 32 summarises the key distinctions. Table 32: Comparison between Kim et al. [2025] and DISCA. Dimension
Kim et al. (2025)
DISCA (this work)
Goal Personas Persona source Metric Correction Models Countries Scenarios Theory
Measure persona sensitivity 7 sociodemographic categories Author-designed binary labels MDD (persona-pair distance) None (diagnostic only) 3 (GPT-4o, GPT-3.5, Llama-2) Aggregate (Western vs. Eastern) 9 dimensions, 10k synthetic Partisan sorting (descriptive)
Correct cultural misalignment 4 WVS-grounded agents per country Empirical WVS-7 microdata MIS (ℓ2 to human AMCE) PT–IS + dual-pass reliability gate 7 checkpoints, 5 families, 2B–70B 20 individual countries 6 dimensions, 310–500 per country Variance-aware MSE-optimal shrinkage (Prop. 1)
Three findings from Kim et al. [2025] directly inform DISCA’s design. Persona sensitivity motivates disagreement-driven steering. Kim et al. show that LLMs exhibit moral-decision distances (MDD) 2–4× larger than humans under contrasting personas. This confirms that persona prompts do shift LLM moral reasoning substantially, validating the premise that within-country persona spread is an informative signal. DISCA converts this sensitivity from a liability (uncontrolled decision shifts) into a feature (a sufficient statistic for correction reliability, Proposition 1). The partisan sorting phenomenon justifies WVS grounding. Kim et al. find that political orientation produces the largest persona effect in LLMs, far exceeding its effect in human respondents. This “partisan sorting” effect reveals that LLMs over-index on politically salient dimensions when given sociodemographic labels. DISCA avoids this pathology by grounding personas in empirical WVS-7 microdata rather than author-designed binary labels: the ten-dimensional cultural profile 47
(Table 26) spreads influence across religiosity, child-rearing values, social trust, and other axes, so no single dimension monopolises the correction. The leave-one-WVS-dim-out probe (Table 28) confirms that influence is spread across multiple WVS axes rather than monopolised by one-individual cells can be large in either direction, but the load-bearing dimensions vary by MultiTP attribute, with no single WVS descriptor dominating across the board. Moral flips validate the reliability gate. Kim et al. document that specific personas can entirely reverse LLM preferences on certain moral dimensions (e.g., social status under a progressive persona in GPT-4o). DISCA’s dual-pass reliability gate (Eq. 5) is designed precisely for this regime: when persona-driven corrections are large enough to flip the decision, the two independent IS passes are likely to disagree, and the exponential decay weight r shrinks the correction toward zero. The tail-safety analysis (Table 15) confirms this mechanism reduces harmed cells from 11/120 to 3/120 compared to ungated consensus. From diagnosis to correction. The fundamental advance of DISCA over the Kim et al. framework is operational: where they diagnose persona sensitivity, we harness it. Their MDD metric measures the distance between two contrasting personas; our MIS measures the distance to the human target and actively reduces it. The variance-aware shrinkage result (Proposition 1) provides the formal bridge: within-panel disagreement (analogous to their MDD) is a sufficient statistic for the correction’s reliability, and the shrinkage estimator converts this diagnostic signal into the MSE-optimal scalar correction.
A17
Extended Limitations
This appendix expands the limitations summarised in §4.2 with technical caveats not covered in the main text. WVS-to-trolley linkage. The mapping from WVS value dimensions to trolley moral dimensions operates through both direct thematic overlap and indirect modulation (App. A12.3). The 10-dimensional WVS profile explains R2 ∈ [0.55, 0.69] of human AMCE variance, and a leaveone-dimension-out probe identifies load-bearing couplings alongside noisy ones that the IS stage automatically down-weights. We avoid claims of the form “WVS dimension X caused AMCE shift Y ,” but the ensemble design and the IS noise-filtering mechanism together ensure the method is robust even when individual WVS–trolley links are imperfect. Inter-persona reward assumption. The reward ri = δi − δbase assumes persona shifts approximate human target directions. This is plausible when the base model represents a WEIRD-biased prior, but if a persona’s shift is orthogonal to the human target, the signal may mislead. Quantifying whether persona-consensus directions correlate with country-level human AMCE vectors independently of the base model-and when corrections diverge from targets-is open; we do not report that correlation matrix here. Geographic coverage, preprocessing, and WVS vintage. We report 20 countries out of 100+ MultiTP countries; results depend on the preprocessing pipeline in App. A14. WVS Wave 7 (∼2017–2022) is a fixed historical slice that may not reflect current preferences in rapidly evolving societies; persona quality is weakest where coverage is sparse (e.g., ETH and BGD in our panel; Saudi Arabia is excluded entirely and uses a manual fallback persona, App. A12.2). No low-resource languages (Amharic, Khmer, Yoruba) appear in the panel; cross-lingual generalisation to such settings remains untested, and one language per country may conflate linguistic and cultural effects. Model quantisation. 4-bit/Int8 quantisation may introduce logit artefacts affecting IS stabilityobserved empirically on Unsloth Phi-4 prior to switching to vLLM BF16 for the headline runs.
48
NeurIPS Paper Checklist The checklist is designed to encourage best practices for responsible machine learning research, addressing issues of reproducibility, transparency, research ethics, and societal impact. Do not remove the checklist: The papers not including the checklist will be desk rejected. The checklist should follow the references and follow the (optional) supplemental material. The checklist does NOT count towards the page limit. Please read the checklist guidelines carefully for information on how to answer these questions. For each question in the checklist: • You should answer [Yes], [No], or [N/A]. • [N/A] means either that the question is Not Applicable for that particular paper or the relevant information is Not Available. • Please provide a short (1–2 sentence) justification right after your answer (even for [N/A]). The checklist answers are an integral part of your paper submission. They are visible to the reviewers, area chairs, senior area chairs, and ethics reviewers. You will also be asked to include it (after eventual revisions) with the final version of your paper, and its final version will be published with the paper. The reviewers of your paper will be asked to use the checklist as one of the factors in their evaluation. While [Yes] is generally preferable to [No], it is perfectly acceptable to answer [No] provided a proper justification is given (e.g., error bars are not reported because it would be too computationally expensive” or “we were unable to find the license for the dataset we used”). In general, answering [No] or [N/A] is not grounds for rejection. While the questions are phrased in a binary way, we acknowledge that the true answer is often more nuanced, so please just use your best judgment and write a justification to elaborate. All supporting evidence can appear either in the main paper or the supplemental material, provided in appendix. If you answer [Yes] to a question, in the justification please point to the section(s) where related material for the question can be found. IMPORTANT, please: • Delete this instruction block, but keep the section heading “NeurIPS Paper Checklist", • Keep the checklist subsection headings, questions/answers and guidelines below. • Do not modify the questions and only use the provided macros for your answers. 1. Claims Question: Do the main claims made in the abstract and introduction accurately reflect the paper’s contributions and scope? Answer: [Yes] Justification: The abstract and §1 state three claims: (i) DISCA is a train-free inferencetime controller that requires no per-country reward models or extra training data; (ii) it cuts macro MIS by 10–24% on seven open-weight backbones across 20 countries and 6 moral dimensions; (iii) calibration competes with scale (Phi-4 14B beats Llama-3.3-70B in absolute MIS). All three are substantiated in Table 2 and Figure 3; scope and assumptions (binary dilemmas with decision-token logits, WVS-7 vintage) are made explicit in §4.2 and Appendix A17. Guidelines: • The answer [N/A] means that the abstract and introduction do not include the claims made in the paper. • The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A [No] or [N/A] answer to this question will not be perceived well by the reviewers. • The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings. • It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper. 49
2. Limitations Question: Does the paper discuss the limitations of the work performed by the authors? Answer: [Yes] Justification: §4.2 bounds generalisation with three caveats (alignment to a survey statistic vs. perceived legitimacy; PT value function as aggregation kernel rather than cognitive model; decision-token logit requirement). Appendix A17 expands these with: WVSto-trolley linkage being indirect, the inter-persona reward assumption, geographic coverage (20 of 100+ MultiTP countries; sparse personas for under-sampled countries), WVS Wave 7 vintage, and 4-bit/Int8 quantisation artefacts on IS stability. Guidelines: • The answer [N/A] means that the paper has no limitation while the answer [No] means that the paper has limitations, but those are not discussed in the paper. • The authors are encouraged to create a separate “Limitations” section in their paper. • The paper should point out any strong assumptions and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be. • The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated. • The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon. • The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size. • If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness. • While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren’t acknowledged in the paper. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community. Reviewers will be specifically instructed to not penalize honesty concerning limitations. 3. Theory assumptions and proofs Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof? Answer: [Yes] Justification: The single theoretical result, Proposition 1 (the MSE-optimal scalar shrinkage of the consensus correction is a closed-form function of within-panel variance), is stated in the main text with all assumptions (i.i.d. persona noise, ℓ2 risk). The full proof is given in Appendix A2, and the connection to Eq. 3 appears in §3. Guidelines: • The answer [N/A] means that the paper does not include theoretical results. • All the theorems, formulas, and proofs in the paper should be numbered and crossreferenced. • All assumptions should be clearly stated or referenced in the statement of any theorems. • The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition. • Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material. 50
• Theorems and Lemmas that the proof relies upon should be properly referenced. 4. Experimental result reproducibility Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)? Answer: [Yes] Justification: §3 gives the full algorithm (Algorithm 1, Eqs. 3–5); the full hyperparameter table (Table 19) lists every numerical default with the validation procedure (synthetic 200-scenario pool, transferability check on a 50-scenario MultiTP English subset). All backbones are publicly available open-weight checkpoints. Preprocessing (cap=80, oversample=36), seeds (3), and per-country preprocessing details are documented in Appendices A14, A9.1. We additionally release a self-contained reference implementation (DISCA_Alignment/) with the same defaults and a single-command sweep entry point. Guidelines: • The answer [N/A] means that the paper does not include experiments. • If the paper includes experiments, a [No] answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not. • If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable. • Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general. releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model (e.g., in the case of a large language model), releasing of a model checkpoint, or other means that are appropriate to the research performed. • While NeurIPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution. For example (a) If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm. (b) If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully. (c) If the contribution is a new model (e.g., a large language model), then there should either be a way to access this model for reproducing the results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset). (d) We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source models, it may be that access to the model is limited in some way (e.g., to registered users), but it should be possible for other researchers to have some path to reproducing or verifying the results. 5. Open access to data and code Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? Answer: [Yes] Justification: We release DISCA_Alignment/ (anonymised for review; MIT-licensed) as supplemental material. The package is a self-contained slice of the research codebase: a single CLI entry point (run.py) drives the per-model 20-country sweep, with backend selection (vllm / hf_native / unsloth) and explicit flags for the three external data paths (–multitp-data, –wvs-data, –human-amce). The README documents install (Python 3.10+, CUDA 12.x), HF token wiring for gated models, exact commands, and the output 51
schema (per-country baseline / DISCA CSVs and a comparison file). MultiTP, WVS-7 Wave, and the human AMCE table are publicly available datasets cited in §4; access instructions are reproduced in the README. Guidelines: • The answer [N/A] means that paper does not include experiments requiring code. • Please see the NeurIPS code and data submission guidelines (https://neurips. cc/public/guides/CodeSubmissionPolicy) for more details. • While we encourage the release of code and data, we understand that this might not be possible, so [No] is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark). • The instructions should contain the exact command and environment needed to run to reproduce the results. See the NeurIPS code and data submission guidelines (https: //neurips.cc/public/guides/CodeSubmissionPolicy) for more details. • The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc. • The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why. • At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable). • Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted. 6. Experimental setting/details Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer) necessary to understand the results? Answer: [Yes] Justification: There is no training (the method is inference-time on frozen weights); we therefore detail all decoding and controller settings. §4 specifies the dataset (MultiTP, 20 countries), preprocessing (deduplication, category cap=80, oversample to ≥36 per dimension), the seven main backbones with sizes, and the metric definitions (MIS, Pearson r, JSD, win rate). All hyperparameters and how they were chosen are tabulated in Table 19 with the synthetic-pool grid-search procedure described in Appendix A11. Sensitivity sweeps for the temperature family (Appendix A11), category cap (Appendix A14), and the controller knobs (s, λcoop , σ, Tcat ) appear in Appendix A11. Guidelines: • The answer [N/A] means that the paper does not include experiments. • The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them. • The full details can be provided either with the code, in appendix, or as supplemental material. 7. Experiment statistical significance Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments? Answer: [Yes] Justification: The headline table (Table 2) reports macro MIS as mean ± standard deviation across 3 independent seeds (independent IS noise draws; the underlying model weights and data are fixed). A separate multi-seed stability analysis (Appendix A9.1) gives per-country variability, and a bootstrap confidence interval on JSD (±0.004, Appendix A11) quantifies the noise floor for distributional metrics. We explicitly state that conclusions are stable well within the bootstrap noise floor; sensitivity sweeps along nine independent axes confirm this. Guidelines: • The answer [N/A] means that the paper does not include experiments. 52
• The authors should answer [Yes] if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper. • The factors of variability that the error bars are capturing should be clearly stated (for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions). • The method for calculating the error bars should be explained (closed form formula, call to a library function, bootstrap, etc.) • The assumptions made should be given (e.g., Normally distributed errors). • It should be clear whether the error bar is the standard deviation or the standard error of the mean. • It is OK to report 1-sigma error bars, but one should state it. The authors should preferably report a 2-sigma error bar than state that they have a 96% CI, if the hypothesis of Normality of errors is not verified. • For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range (e.g., negative error rates). • If error bars are reported in tables or plots, the authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text. 8. Experiments compute resources Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? Answer: [Yes] Justification: Appendix A15 (Table 31) reports per-scenario wall-clock on a single H100 GPU: 0.083–0.086 s/scenario for DISCA at K ∈ {64, 128, 256} vs. 0.023 s/scenario for vanilla decoding (3.57–3.72× overhead). Backbone-level cost is given in Figure 5: 350 ms/scenario for Phi-4 vs. 1414 ms for Llama-3.3-70B with DISCA. Sub-70B backbones run on a single H100 in BF16 via vLLM; the 70B backbone uses 4-bit Unsloth on the same hardware. The released code’s CLI also runs end-to-end on a Kaggle T4/P100 dual-GPU node for sub-14B models. Exploratory experiments on a small 3-backbone × 5-country prototyping panel consumed additional compute beyond the reported runs. Guidelines: • The answer [N/A] means that the paper does not include experiments. • The paper should indicate the type of compute workers CPU or GPU, internal cluster, or cloud provider, including relevant memory and storage. • The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute. • The paper should disclose whether the full research project required more compute than the experiments reported in the paper (e.g., preliminary or failed experiments that didn’t make it into the paper). 9. Code of ethics Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines? Answer: [Yes] Justification: The authors have reviewed the NeurIPS Code of Ethics. The work uses only public, ethically-collected datasets (MultiTP [Jin et al., 2025], World Values Survey Wave 7, Moral Machine human AMCEs [Awad et al., 2018]); it conducts no new humansubjects research and introduces no scraped data. Anonymity is preserved in the submission. Negative-impact considerations (risk of encoding harmful majorities) are explicitly discussed in §4.2 with a per-persona utility floor mitigation validated in Appendix A11 (Table 25). Guidelines: • The answer [N/A] means that the authors have not reviewed the NeurIPS Code of Ethics. 53
• If the authors answer [No], they should explain the special circumstances that require a deviation from the Code of Ethics. • The authors should make sure to preserve anonymity (e.g., if there is a special consideration due to laws or regulations in their jurisdiction). 10. Broader impacts Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed? Answer: [Yes] Justification: §4.2 discusses both directions. Positive: a train-free, low-overhead controller lets practitioners respect cross-national preference heterogeneity at deployment time without finetuning per country, lowering the barrier to culturally pluralistic alignment. Negative: aligning to majority survey statistics could entrench harmful majorities or marginalise minority views [Zewail et al., 2026], and alignment to a survey statistic is not equivalent to legitimacy [Atari et al., 2023]. We propose and validate a per-persona utility floor that caps how far any single persona’s post-correction utility can drop from vanilla (Appendix A11, Table 25), and we are explicit that DISCA is a research artefact, not a deployment-ready normative system. Guidelines: • The answer [N/A] means that there is no societal impact of the work performed. • If the authors answer [N/A] or [No], they should explain why their work has no societal impact or why the paper does not address societal impact. • Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (e.g., deployment of technologies that could make decisions that unfairly impact specific groups), privacy considerations, and security considerations. • The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to generate Deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster. • The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from (intentional or unintentional) misuse of the technology. • If there are negative societal impacts, the authors could also discuss possible mitigation strategies (e.g., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML). 11. Safeguards Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pre-trained language models, image generators, or scraped datasets)? Answer: [N/A] Justification: We do not release any pretrained models, generative models, or scraped datasets. The released artefact (DISCA_Alignment/) is an inference-time controller layered over publicly available open-weight LLMs whose existing safeguards remain in place. The method itself produces no novel generative capability beyond that of the underlying models; the per-persona utility floor (Appendix A11, Table 25) is included as an in-method safeguard against majoritarian skew. Guidelines: • The answer [N/A] means that the paper poses no such risks. 54
• Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters. • Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images. • We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort. 12. Licenses for existing assets Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected? Answer: [Yes] Justification: All third-party assets are cited with version information. Datasets: MultiTP [Jin et al., 2025], the Moral Machine human AMCE table [Awad et al., 2018], and World Values Survey Wave 7 (inverted CSV v6_0) are used under the terms of the respective public releases. Models: Phi-4 and Phi-3.5-mini (MIT), Llama-3.3-70B (Llama-3 Community License), Qwen-2.5 / Qwen3-VL (Apache-2.0 / Qwen license), Gemma family (Gemma terms of use), Magistral-Small-2509 (Apache-2.0). Code dependencies (Transformers, vLLM, Unsloth, bitsandbytes) are listed with version pins in the released requirements.txt; we use them under their stated licenses (mostly Apache-2.0 / MIT). The compute_ACME reference implementation we follow is credited to Jin et al. [2025] in Appendix A13. Guidelines: • The answer [N/A] means that the paper does not use existing assets. • The authors should cite the original paper that produced the code package or dataset. • The authors should state which version of the asset is used and, if possible, include a URL. • The name of the license (e.g., CC-BY 4.0) should be included for each asset. • For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided. • If assets are released, the license, copyright information, and terms of use in the package should be provided. For popular datasets, paperswithcode.com/datasets has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset. • For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided. • If this information is not available online, the authors are encouraged to reach out to the asset’s creators. 13. New assets Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets? Answer: [Yes] Justification: We release the DISCA reference implementation (DISCA_Alignment/) as a new code asset under the MIT License. The package ships with: a top-level README documenting installation, backend selection (vLLM / HF native / Unsloth), the three external data flags, environment variables, the quick-start command, and the output schema; a flat disca/ module with one-purpose files (controller, dual-pass core, runner, AMCE/metrics, persona generation, multilingual scenarios, model loader); pinned dependencies in requirements.txt; and a single CLI entry point (run.py –help). The submission archive is anonymised. All hyperparameter defaults match Table 19 in the main paper. Guidelines: 55
• The answer [N/A] means that the paper does not release new assets. • Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc. • The paper should discuss whether and how consent was obtained from people whose asset is used. • At submission time, remember to anonymize your assets (if applicable). You can either create an anonymized URL or include an anonymized zip file. 14. Crowdsourcing and research with human subjects Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)? Answer: [N/A] Justification: The paper conducts no new crowdsourcing or human-subjects research. Human reference data is reused from the existing public Moral Machine experiment [Awad et al., 2018] as packaged by MultiTP [Jin et al., 2025], and from the World Values Survey Wave 7 public microdata; we make no new contact with participants and recruit no annotators. Guidelines: • The answer [N/A] means that the paper does not involve crowdsourcing nor research with human subjects. • Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper. • According to the NeurIPS Code of Ethics, workers involved in data collection, curation, or other labor should be paid at least the minimum wage in the country of the data collector. 15. Institutional review board (IRB) approvals or equivalent for research with human subjects Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained? Answer: [N/A] Justification: No new human-subjects research is conducted. The work is a secondary analysis of pre-existing public datasets (Moral Machine human AMCEs, MultiTP scenarios, WVS-7 microdata), each of which carries its own ethics review at the original collection point; no further IRB approval applies. Guidelines: • The answer [N/A] means that the paper does not involve crowdsourcing nor research with human subjects. • Depending on the country in which research is conducted, IRB approval (or equivalent) may be required for any human subjects research. If you obtained IRB approval, you should clearly state this in the paper. • We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the guidelines for their institution. • For initial submissions, do not include any information that would break anonymity (if applicable), such as the institution conducting the review. 16. Declaration of LLM usage Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the core methodology, scientific rigor, or originality of the research, declaration is not required. 56
Answer: [Yes] Justification: LLMs are central to the work in two ways, both fully described. (i) Frozen open-weight LLMs (Phi-4, Llama-3.3-70B, Magistral-24B, Qwen3-VL-8B, Qwen2.5-7B, Phi-3.5-mini, Gemma-4-E2B) are the subject that DISCA steers; their identities, sizes, and decoding settings are documented in §4 and Appendix A7. (ii) For the open-ended extension (§3.4), an LLM parses free-form responses into a (choice, confidence) pair to recover a pseudo-logit-gap; this auxiliary use is documented in that section. A separate synthetic 200-scenario tuning pool generated by GPT-4 is disclosed in Appendix A11 along with its limitations. LLMs were not used to generate the paper’s claims, results, or proofs. Guidelines: • The answer [N/A] means that the core method development in this research does not involve LLMs as any important, original, or non-standard components. • Please refer to our LLM policy in the NeurIPS handbook for what should or should not be described.
57