arXiv:2604.11141v1 [cs.LG] 13 Apr 2026
Reducing Hallucination in Enterprise AI Workflows via Hybrid Utility Minimum Bayes Risk (HUMBR) Chenhao Fang∗
Jordi Mola∗
Mark Harman
Meta Platforms, Inc. Menlo Park, CA, USA [email protected]
Meta Platforms, Inc. Bellevue, WA, USA [email protected]
Meta Platforms, Inc. London, UK [email protected]
Jason Nawrocki
Vaibhav Shrivastava
Yue Cheng
Meta Platforms, Inc. Andover, MA, USA [email protected]
Meta Platforms, Inc. Bellevue, WA, USA [email protected]
Meta Platforms, Inc. Bellevue, WA, USA [email protected]
Jay Shah
Katayoun Zand
Mansi Tripathi
Meta Platforms, Inc. New York, NY, USA [email protected]
Meta Platforms, Inc. Menlo Park, CA, USA [email protected]
Meta Platforms, Inc. Bellevue, WA, USA [email protected]
Arya Pudota
Matthew Becker
Hervé Robert
Meta Platforms, Inc. Bellevue, WA, USA [email protected]
Meta Platforms, Inc. Bellevue, WA, USA [email protected]
Meta Platforms, Inc. Menlo Park, CA, USA [email protected]
Abhishek Gulati Meta Platforms, Inc. Menlo Park, CA, USA [email protected]
Abstract Although LLMs drive automation, it is important to ensure immense consideration for high-stakes enterprise workflows such as those involving legal matters, risk management, and privacy compliance. For Meta, and other organizations like ours, a single hallucinated clause in such high stakes workflows may introduce greater risks. We show that by framing hallucination mitigation as a Minimum Bayes Risk (MBR) problem, we can reduce this risk. Specifically, we introduce a Hybrid Utility MBR (HUMBR) framework that synthesizes semantic embedding similarity with lexical precision to identify consensus without ground-truth references, for which we derive rigorous error bounds. We complement this theoretical analysis with a comprehensive empirical evaluation on widely-used public benchmark suites
(TruthfulQA and LegalBench) and also real world data from Meta production deployment. The results from our empirical study show that MBR significantly outperforms standard Universal Self-Consistency. Notably, 81% of the pipeline’s suggestions were preferred over human-crafted ground truth, and critical recall failures were also reduced.
CCS Concepts • Computing methodologies → Artificial intelligence.
Keywords Large Language Models, Minimum Bayes Risk, Hallucination Mitigation, Ensemble Learning
∗ Authors contributed equally.
Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for components of this work owned by others than ACM must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. Request permissions from [email protected]. KDD ’26, Jeju, South Korea © 2026 ACM. ACM ISBN 978-x-xxxx-xxxx-x/YYYY/MM https://doi.org/10.1145/XXXXXXX.XXXXXXX
ACM Reference Format: Chenhao Fang, Jordi Mola, Mark Harman, Jason Nawrocki, Vaibhav Shrivastava, Yue Cheng, Jay Shah, Katayoun Zand, Mansi Tripathi, Arya Pudota, Matthew Becker, Hervé Robert, and Abhishek Gulati. 2026. Reducing Hallucination in Enterprise AI Workflows via Hybrid Utility Minimum Bayes Risk (HUMBR). In Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD ’26), August 9– 13, 2026, Jeju, South Korea. ACM, New York, NY, USA, 11 pages. https: //doi.org/10.1145/XXXXXXX.XXXXXXX
KDD ’26, August 9–13, 2026, Jeju, South Korea
1
Introduction
Many organizations, including Meta, are using Large Language Models (LLMs) to catalyze a paradigm shift in the way information is processed: moving from rigid, rule-based automation to flexible, semantic-aware intelligence [7, 9, 28]. However, there is a well-known trust gap [12] due to the inherent stochastic nature of generative AI. In particular, AI hallucination presents a severe existential risk when AI is applied to high-stakes corporate workflows such as legal discovery, privacy compliance engineering, and regulatory interpretation. The challenge is that single-model generations are inherently unstable; even State-Of-The-Art models can confidently fabricate details when operating on long-tail knowledge or complex logical predicates [27]. Standard industrial approaches to mitigate hallucination typically rely on Retrieval-Augmented Generation [21] or iterative refinement [4]; ‘Ask the model to critique itself’. Unfortunately, these methods have distinct failure modes in production. Relying on a single decoding path even with temperature 𝑇 = 0 leaves the system vulnerable to the specific biases and blind spots of that model instance. Another common ensemble technique is to generate multiple responses and ask an LLM to summarize them. But this is often flawed in high-precision tasks. Summarization introduces a second order of generation, creating a risk of compounding hallucinations where the summarizer conflates conflicting details or smooths over nuances needed for compliance [25]. To address the challenge of minimizing hallucination risk for high-stakes enterprise workflows, we draw inspiration from decision theory, introducing the Minimum Bayes Risk (MBR) approach to hallucination mitigation. Our key insight is that, while individual models may hallucinate, they tend to hallucinate differently [4, 20]. True information acts as an attractor in the semantic space—diverse models and decoding paths tend to converge on the truth—whereas hallucinations are often stochastic outliers [26]. By generating an ensemble of candidates across heterogeneous models and temperatures, and selecting the centroid candidate that maximizes the utility function, we can mathematically filter out noise without needing a ground-truth reference. Our primary contributions are to: • introduce Hybrid Utility MBR (HUMBR), a robust algorithm for LLM answer selection. • prove that HUMBR asymptotically approaches ground truth under reasonable sparse hallucination assumptions. • show that HUMBR can be optimally configured to satisfy a strict hallucination tolerance. • report results of a comprehensive empirical study using public benchmarks and real-world deployment. The results reveal that HUMBR outperforms with statistical significance (general and domain-specific) single-model baselines and standard ensemble method and also human experts.
2 Related Work 2.1 Hallucination Mitigation in LLMs “Hallucination”, the generation of plausible but factually incorrect content, remains a major barrier to enterprise LLM adoption [12],
Fang and Mola, et al.
typically categorized into intrinsic (contradicting inputs) and extrinsic (fabricating details) types [27]. Standard mitigation strategies operate at different stages of the model lifecycle: RLHF [22] aligns models during training, while Retrieval-Augmented Generation (RAG) [15] grounds inference in external evidence. However, generators may still fail due to attention lapses or context loss [18]. Consequently, post-hoc verification has become critical, with approaches like SelfCheckGPT [20] exploiting the stochastic instability of hallucinations to detect inconsistencies.
2.2
Advanced Decoding and Inference-Time Interventions
Recent research manipulates the decoding process to enhance factuality. Techniques like Inference-Time Intervention (ITI) [16] and DoLa [2] shift activation heads or contrast layers to amplify truthful signals. Similarly, Context-Aware Decoding (CAD) [23] penalizes prior probabilities to enforce reliance on provided context. While effective, these token-level interventions often require task-specific tuning. Our HUMBR approach offers a robust, sequence-level alternative that optimizes for holistic consensus rather than local token probabilities.
2.3
Ensemble Learning and Consensus
Ensemble methods leverage diversity to improve robustness. SelfConsistency [26] demonstrated that majority voting boosts chainof-thought performance, though naive voting is brittle for openended generation. To address this, Fang et al. [8] proposed LLMEnsemble, a probabilistic framework for weighted consensus that treats generation as a truth discovery problem. Alternative consensus strategies include Multi-Agent Debate [5, 17], which induces agreement through dialogue, and Generatethen-Rank pipelines like LLM-Blender [13] or trained Verifiers [3]. However, these often suffer from high latency or require domainspecific training. Our approach distinguishes itself as reference-free, unsupervised, and single-turn, using embedding similarity to find a “centroid” consensus without auxiliary models [1].
2.4
Minimum Bayes Risk (MBR) Decoding
Originally from Statistical Machine Translation [11], MBR has replaced MAP decoding in contexts where probability concentration leads to repetitive outputs [6]. Recently readapted for LLMs, MBR has been applied to instruction following [24] and translation [10] using metrics like BERTScore or COMET. While Kuhn et al. [14] explored similar “semantic uncertainty,” existing works typically rely on single metrics. We advance this by introducing a Hybrid Utility Function, combining semantic embeddings with lexical ROUGE scores to ensure both conceptual accuracy and structural coherence.
3
Methodology
Standard autoregressive decoding implicitly assumes that the mode of the distribution corresponds to factual truth. However, in highentropy open-domain generation, this assumption frequently fails as probability mass often concentrates on generic or hallucinated sequences. To address this, we reframe hallucination mitigation not as a likelihood maximization task, but as a Minimum Bayes Risk (MBR) decision problem. Our core hypothesis is geometric: while
Reducing Hallucination in Enterprise AI Workflows via Hybrid Utility Minimum Bayes Risk (HUMBR)
Phase 1: Generation
KDD ’26, August 9–13, 2026, Jeju, South Korea
Phase 2: Consensus Calculation
Phase 3: Selection
Pairwise Hybrid Utility Function
𝑈𝑖 𝑗 = 𝛼 · CosSim (𝐸𝑖 , 𝐸 𝑗 ) + (1 − 𝛼 ) · ROUGE (𝐶𝑖 , 𝐶 𝑗 )
𝐶1
1.00
0.85
...
0.88
0.86
𝐶2
0.91
1.00
...
0.94
0.95
ng
S
Prompt (Input)
pli am
LLMs
Check 𝑆𝑚𝑎𝑥 > 𝜏 ?
PASS
𝐶2
.. .
𝑇
.. .
< 0.7
ABSTAIN
𝐶𝑁
0.40
0.84
𝐶1
𝐶2
...
1.00
0.75
𝐶𝑁
Score 𝑆𝑖
HUMBR Consensus Framework
Figure 1: Workflow of the proposed HUMBR ensemble system. The system calculates consensus using a Hybrid Utility Function (Semantic + Lexical) and applies a consensus threshold 𝜏 to filter risky outputs. hallucinations are stochastically dispersed, factual answers tend to cluster in a shared semantic neighborhood. Therefore, the truth manifests as the centroid of the semantic probability mass, rather than the sharpest peak of the density function. The overall architecture of our system is illustrated in Figure 1. The workflow proceeds in three phases: diverse sampling to approximate the posterior, pairwise consensus calculation, and risk-aware selection. We structure this section as follows: • We first establish the theoretical foundations of referencefree MBR in Section 3.1. • Moving from theory to reality, Section 3.2 derives a generalized error bound that explicitly accounts for intra-model correlations. • Leveraging these bounds, Section 3.3 solves the engineering optimization problem, determining the minimal ensemble size required to meet a strict quality threshold. • Finally, Section 3.4 details the practical implementation, introducing our Hybrid Utility Function and the complete HUMBR algorithm.
3.1
Minimum Bayes Risk Formulation
Let 𝑃 (𝑦|𝑥) denote the true posterior distribution of valid responses 𝑦 given an input prompt 𝑥. In generative tasks, our objective is to select a hypothesis 𝑦ˆ from a hypothesis space H that minimizes the expected loss relative to the true distribution. ˆ :Y× Definition 3.1 (Bayes Risk). Given a loss function L (𝑦, 𝑦) Y → R ≥0 which quantifies the divergence between a reference 𝑦 ˆ the Bayes Risk 𝑅(𝑦|𝑥) ˆ is defined as the expected and a hypothesis 𝑦,
loss under the posterior: ˆ = E𝑦∼𝑃 (· |𝑥 ) [L (𝑦, 𝑦)] ˆ = 𝑅(𝑦|𝑥)
∑︁
ˆ 𝑃 (𝑦|𝑥)L(𝑦, 𝑦)
(1)
𝑦∈Y
The optimal selection strategy, known as MBR selection, selects the candidate that minimizes this risk: ˆ 𝑦ˆMBR = arg min 𝑅(𝑦|𝑥)
(2)
𝑦ˆ ∈ H
In reference-free settings where the true posterior 𝑃 (𝑦|𝑥) is intractable, we approximate the expectation using Monte Carlo integration. We assume the ensemble of generated candidates C = {𝑐 1, . . . , 𝑐 𝑁 } are independent samples drawn from the model’s apˆ = proximate posterior 𝑃𝜃 (𝑦|𝑥). By defining a utility function U (𝑦, 𝑦) ˆ minimizing risk becomes equivalent to maximizing ex1 − L (𝑦, 𝑦), pected utility: 1 ∑︁ ˆ 𝑦ˆMBR ≈ arg max U (𝑐𝑖 , 𝑦) (3) 𝑁 𝑐 ∈C 𝑦ˆ ∈ C 𝑖
Mathematically, 𝑦ˆMBR maximizes the lower bound of the expected utility under the true distribution, thus selects the optimal candidate. The full proof is provided in Appendix A.
3.2
Risk Bounds under Intra-Model Correlation
While MBR maximizes expected utility, practical deployment in zero-tolerance domains requires a rigorous quality guarantee. Specifically, we must bound the probability that the system confidently outputs a hallucination. A limitation of prior analysis is the assumption of independence. In an ensemble of 𝐾 models generating 𝑀 samples each (total
KDD ’26, August 9–13, 2026, Jeju, South Korea
𝑁 = 𝐾𝑀), samples from the same model exhibit strong correlations (𝜌). We derive a generalized error bound that explicitly accounts for this correlation using a hierarchical framework. Let 𝑍𝑘 denote the effective count of divergent (hallucinated) samples produced by model 𝐿𝑘 . While the utility space is continuous, for the purpose of a worst-case risk analysis, we discretize the outcome into binary states: a sample is either consistent with the consensus mode or divergent. To capture the intra-model dependency, we model the divergent count 𝑍𝑘 via a hierarchical Beta-Binomial process: 𝜋𝑘 ∼ Beta(𝑎, 𝑏) (4) 𝑍𝑘 |𝜋𝑘 ∼ Binomial(𝑀, 𝜋𝑘 ) where 𝜋𝑘 is the latent error rate of model 𝑘. The marginal distribution of 𝑍𝑘 is governed by the mean error rate 𝜇 and the correlation coefficient 𝜌: 𝐵(𝑧 + 𝜇 ( 1 − 1), 𝑀 − 𝑧 + (1 − 𝜇)( 1 − 1)) 𝑀 𝜌 𝜌 𝑃 (𝑍𝑘 = 𝑧) = (5) 𝑧 𝐵(𝜇 ( 𝜌1 − 1), (1 − 𝜇)( 𝜌1 − 1)) We introduce a global Consensus Threshold 𝜏 ∈ [0.5, 1]. The system is considered valid only if the consensus score 𝑆 exceeds 𝜏. In the worst-case analysis, a failure occurs when the hallucinations cluster so densely that their aggregated mass exceeds this threshold, tricking the MBR algorithm into selecting a falsehood. Let z = (𝑧 1, . . . , 𝑧𝐾 ) be the vector of divergent counts across 𝐾 models. The Failure Region Ω F (𝜏) is defined as the state space where the total divergent mass constitutes a super-majority: ) ( Í𝐾 𝑘=1 𝑧𝑘 𝐾 ≥𝜏 (6) Ω F (𝜏) = z ∈ Z ≥0 𝑁 Theorem 3.2 (Threshold-Aware Risk Bound). For an ensemble operating under threshold 𝜏, the probability of a failure is the cumulative probability mass of the Beta-Binomial product distribution over Ω F (𝜏): " 𝐾 # ∑︁ Ö 𝑃 fail = 𝑃 BetaBinomial (𝑧𝑘 ; 𝑀, 𝜇𝑘 , 𝜌𝑘 ) (7) z∈Ω F (𝜏 )
𝑘=1
By increasing 𝜏 (e.g., from 0.5 to 0.7), we shrink the volume of Ω F , exponentially reducing the risk of accepting a hallucination, albeit at the cost of abstaining on ambiguous queries.
3.3
Enterprise System Design: Meeting Quality Constraint
Armed with the generalized error bound derived in Theorem 3.2, we address the inverse engineering problem: How to configure the ensemble to guarantee a specific error bound 𝜖 with minimum computational cost? 3.3.1 Optimization Formulation. We formulate this as a constrained optimization problem. Let 𝐶𝑘 be the inference cost per token for model 𝑘, our objective is to find the optimal subset of models K and generation count 𝑀 such that the failure probability remains below an error tolerance 𝜖 (e.g., 𝜖 = 10−4 ): ∑︁ min 𝐶𝑘 · 𝑀 K,𝑀 (8) 𝑘∈K s.t. 𝑃 fail (Ω F (𝜏); 𝑀, 𝜇𝑘 , 𝜌𝑘 ) ≤ 𝜖
Fang and Mola, et al.
Failure Probability (𝑃 fail ) Baseline (Single Model)
Summarizer
Failure Tolerance 𝜖
Target Zone MBR (Ours) Inference Cost
Figure 2: Cost-𝑃fail Pareto Frontier: An idealized depiction of the operational trade-off landscape.
To intuitively understand the operational landscape defined above, we visualize the idealized Cost-𝑃fail Pareto Frontier in Figure 2. The plot reveals distinct regimes: (1) High Failure Rate Zone: Single-model baselines (lower left) incur minimal computational cost but suffer from stochastic hallucinations (𝑃fail ≫ 𝜖), making them unsuitable for compliance tasks. (2) Inefficient Zone: Naive ensemble methods like "Summarization" increase computational cost linearly but often fail to reduce the failure probability proportionally due to compounding errors. (3) Feasible Zone: Our MBR approach (green centroid) effectively navigates this trade-off. By leveraging the diversity of the ensemble (𝜌 < 1), it pushes the system’s hallucination probability exponentially downward into the Target Zone (satisfying the constraint 𝑃fail ≤ 𝜖) while maintaining a tractable inference budget. 3.3.2 Derivation of Required Sample Size. To solve this efficiently without brute-force simulation, we derive a closed-form approximation for the required number of generations 𝑀. While the ensemble contains heterogeneous models, for the purpose of deriving a conservative design bound, we model the system using effective average parameters. We assume an ensemble of 𝐾 models with ¯ an average error rate 𝜇¯ and an effective pairwise correlation 𝜌. The variance of the hallucination ratio 𝑝ˆ = 𝑍 /𝑁 is inflated by the variance inflation factor (VIF): ¯ 𝜇¯ (1 − 𝜇) ¯ · [1 + (𝑀 − 1) 𝜌] (9) 𝑁 Thus, the ensemble behaves as if it has an effective sample size ˆ = Var(𝑝)
of: 𝐾 ·𝑀 (10) 1 + (𝑀 − 1) 𝜌¯ To ensure the hallucination rate stays below the error tolerance 𝜖, we apply a Hoeffding-type bound. We require the probability that the observed hallucination rate 𝑝ˆ exceeds the consensus threshold 𝜏 to be bounded by 𝜖: ¯ 2 ≤𝜖 𝑃fail ≤ exp −2𝑁𝑒 𝑓 𝑓 (𝜏 − 𝜇) (11) 𝑁𝑒 𝑓 𝑓 =
Reducing Hallucination in Enterprise AI Workflows via Hybrid Utility Minimum Bayes Risk (HUMBR)
where 𝜇¯ is the weighted average error rate of the ensemble. Solving for 𝑀, we obtain the design inequality: 𝑀≥
¯ ln(1/𝜖)(1 − 𝜌) ¯ 2 − 𝜌¯ ln(1/𝜖) 2𝐾 (𝜏 − 𝜇)
(12)
3.3.3 Stratification and Diversity Strategy. Equation 10 reveals operational insights regarding the effective sample size 𝑁𝑒 𝑓 𝑓 : • Independence (𝜌¯ → 0): 𝑁𝑒 𝑓 𝑓 = 𝐾𝑀, the full ensemble contributes. • Collapse (𝜌¯ → 1): 𝑁𝑒 𝑓 𝑓 → 𝐾, adding more samples per model yields zero gain. ¯ proving that model • Saturation: As 𝑀 → ∞, 𝑁𝑒 𝑓 𝑓 → 𝐾/𝜌, diversity (increasing 𝐾) is essential. Illustrative Example. Consider 𝐾 = 4 models, 𝑀 = 4 samples each, with base error rate 𝜇¯ = 0.1 and consensus threshold 𝜏 = 0.7. Under independence (𝜌¯ = 0), the failure probability is ≈ 10−8 . However, with realistic correlation 𝜌¯ = 0.5, the risk rises to ≈ 10−4 — a gap of four orders of magnitude that naive analysis would miss. The intra-model correlation 𝜌¯ is strictly dependent on the sampling strategy. Repeated sampling at a single low temperature yields high correlation (𝜌¯ → 1), while sampling at very high temperatures ¯ degrades quality (𝜇). To optimally balance this trade-off, we propose a Temperature Stratification strategy. Instead of fixing a single temperature 𝑇 , we enforce diversity by sampling along a temperature gradient. For example, in our internal deployment (Sec. 5), to achieve an error tolerance of 𝜖 = 10−4 , our optimization dictated a heterogeneous mix of 𝐾 = 4 models. Rather than generating 𝑀 i.i.d. samples, each model generates exactly one sample at four distinct temperature levels: T = {0, 0.25, 0.5, 0.75} (i.e., 𝑀 = 4). This design forces decor¯ while retaining relation within each model’s outputs (minimizing 𝜌) ¯ high-probability modes at lower temperatures (maintaining low 𝜇).
3.4
Hybrid Utility MBR (HUMBR)
With the ensemble generation protocol established to maximize diversity, we obtain a candidate set C that is robust against correlated hallucinations. The final operational step is to identify the consensus centroid within this pool. We implement the MBR selection mechanism using a Hybrid Utility Function. This function is designed to measure consensus not just by exact string matching which is too brittle for diverse temperatures, but by capturing both conceptual meaning and structural phrasing. We define the utility between candidates 𝑐𝑖 , 𝑐 𝑗 as: U (𝑐𝑖 , 𝑐 𝑗 ) = 𝛼 ·CosSim(𝜙 (𝑐𝑖 ), 𝜙 (𝑐 𝑗 ))+(1−𝛼)·ROUGE-L(𝑐𝑖 , 𝑐 𝑗 ) (13) where 𝜙 (·) is a sentence embedding model and 𝛼 is a weighting hyperparameter (typically 0.6). The complete selection procedure is formalized in Algorithm 1. Note that we introduce the threshold 𝜏. Unlike standard decoding which forces an output, our HUMBR formulation includes an abstention mechanism: if the maximum consensus score 𝑆𝑘 ∗ falls below 𝜏, the system flags the query as ambiguous rather than risking a hallucination, thus enables a controllable trade-off between quality (precision) and coverage (recall). By demanding a super-majority (e.g., 𝜏 = 0.75), we can effectively reject ambiguous cases.
KDD ’26, August 9–13, 2026, Jeju, South Korea
In production, the HUMBR process naturally yields metadata for monitoring model health. By tracking the Divergence Score (average distance from the consensus centroid) for each model in production, we can dynamically estimate the operational error rate 𝜇ˆ𝑘 and selfcorrelation 𝜌ˆ𝑘 . Models that consistently deviate from the ensemble consensus are flagged for retraining or down-weighting in the voting logic.
4
Offline Experiments
To rigorously evaluate the proposed HUMBR framework, we conducted experiments to answer two Research Questions (RQs), identifying whether our method can discern truth from plausible mimicry: • RQ1 Mitigation of Imitative Falsehoods: Can HUMBR effectively filter out imitative falsehoods (common misconceptions) in open-domain scenarios? • RQ2 Domain Precision & Robustness: In high-stakes legal reasoning, does HUMBR outperform standard decoding strategies?
4.1
Experimental Setup
Datasets. We utilize two distinct benchmarks to test breadth and depth: (1) TruthfulQA (Generation Track): A benchmark comprising 817 questions designed to elicit imitative falsehoods. We focus on the Generation task to test open-ended robustness. (2) LegalBench: A legal reasoning benchmark. We focus on two representative task categories that challenge LLM reasoning: • Interpretation (Classification): Five difficult tasks requiring the classification of contractual clauses (e.g., contract_nli). • Rule-Application (Extraction/Generation): Two tasks applying specific rules to facts (e.g., hearsay, rule_recall). Baselines. We focus on reference-free, black-box methods that are deployable via API access. Token-level interventions such as DoLa [2] and ITI [16] require access to model internals (logits or activations), making them inapplicable to proprietary model APIs. We compare our approach against: • Greedy Decoding: Standard generation (𝑇 = 0). • Universal Self-Consistency (USC) [1]: The industry standard for reasoning tasks which leverages LLMs themselves to select the most consistent answer among multiple candidates. • Oracle (Best-of-N): A “cheating” upper bound that has access to the gold reference at selection time. It selects the candidate from the pool C that best matches the ground truth. To avoid circular evaluation—where MBR’s own utility metric would trivially favor itself—the Oracle employs a distinct similarity metric: For classification tasks, we use exact label matching after normalizing model outputs. For generation tasks, we use character-level sequence matching, which computes the ratio of matching subsequences. This Oracle represents the maximum achievable performance given the candidate pool quality. Metrics. We employ a dual-metric strategy to ensure both semantic flexibility and domain rigor:
KDD ’26, August 9–13, 2026, Jeju, South Korea
Fang and Mola, et al.
• Open-Domain (LLM-as-a-Judge): For TruthfulQA, we use Claude-Opus-4.5 to evaluate % Truth × Info and a scalar Semantic Quality Score (1-100). To ensure evaluation robustness, we validated the LLM judge against both itself and human annotations. Details are provided in Appendix D. • Legal Domain (Exact Metrics): For LegalBench, we report Balanced Accuracy for classification tasks and F1 Score for generative rule application, measuring the precision of legal terminology usage. Experimental Setup. We generated a candidate pool of size 𝑁 = 8 per query using a stratified heterogeneous ensemble comprising equal splits from Llama-3.3-70B and Gemini-2.0-Flash. Sampling was conducted via stochastic decoding at temperature 𝑇 = 0.7 to maximize semantic entropy. For the HUMBR consensus mechanism, we set the hybrid utility weight 𝛼 = 0.6, utilizing the opensource DRAMA [19] model for semantic embedding computation. While our production system (Section 5) employs a strict consensus threshold (𝜏 = 0.8) to filter hallucinations, for these offline benchmark experiments, we relaxed the constraint to 𝜏 = 0. This mandates 100% coverage, ensures a rigorous, unconstrained comparison against baselines, which typically lack intrinsic refusal mechanisms. Full details regarding prompt templates and judge specifications are provided in Appendix C.
4.2
RQ1: Mitigation of Imitative Falsehoods
To answer whether geometric consensus can distinguish factual truth from popular misconceptions, we evaluate performance on the TruthfulQA (Generation) benchmark. Table 1: Performance on TruthfulQA. Evaluated by ClaudeOpus-4.5 Judge. % Truth × Info is the strict metric requiring answers to be both factually correct and informative (nonrefusals). Decoding Method
% Truth × Info
Quality Score
Greedy Decoding (𝑇 = 0) Universal Self-Consistency (USC)
69.5% 76.5%
74.1 75.5
HUMBR (Ours) Oracle (Upper Bound)
80.3% 81.5%
81.8 83.5
Results & Analysis. Table 1 presents the comparative results. Our HUMBR approach achieves a +3.8% absolute improvement over the Universal Self-Consistency (USC) baseline. Notably, our method recovers approximately 98.5% of the Oracle performance (80.3% vs 81.5%), suggesting that the selection mechanism is near-optimal given the candidate pool.
4.3
RQ2: Precision in Legal Domain
Legal reasoning tolerates low hallucinations. A model must not only retrieve the correct rule but apply it using precise terms. In this section, we evaluate performance on the Interpretation and Rule-Application subsets of LegalBench. Results. Table 2 compares our HUMBR against standard decoding and Universal Self-Consistency (USC).
Table 2: LegalBench Performance. HUMBR outperforms baselines on both classification (Interpretation) and generation (Rule-Application) tasks. Interpretation
Rule-Application
Method
Acc
Bal-Acc
Exact Match
F1 Score
Greedy Decoding Self-Consistency (USC)
35.1% 39.2%
34.0% 36.5%
44.4% 49.3%
42.9% 45.5%
HUMBR (Ours) Oracle (Upper Bound)
54.2% 69.7%
50.2% 67.1%
53.5% 59.0%
50.8% 53.9%
Analysis. We observe that HUMBR achieves a +15.0% gain in Balanced Accuracy over USC. In legal classification tasks (e.g., determining if a clause is void), USC often suffers from "modal collapse" where the model confidently predicts the majority class due to prior bias. MBR, by weighing the semantic centrality of the generated reasoning paths, effectively filters out these high-confidence but shallow biases. Furthermore, in Rule-Application, the boost in F1 Score (+7.9%) indicates that HUMBR preserves the precise legal wording better. While USC averages out the noise, HUMBR actively selects the candidate that best represents the consensus logic, reducing the risk of omitting key legal qualifiers.
5
Online Deployment: Automated Regulatory Understanding Pipeline at Meta
Following the robust performance of HUMBR on the offline benchmarks, we advanced the system to a live production deployment. We deployed the framework within a regulatory understanding workflow at Meta. This workflow involves parsing complex legal statutes and generating precise, interpretable specifications for engineering teams. Unlike general-purpose chat, this domain requires strict adherence to source texts; a single “hallucinated” obligation or missed exemption can lead to compliance risks.
5.1
Operational Workflow: AI as a Drafting Partner
The system is deployed not as an autonomous agent, but as a highprecision suggestion engine designed to augment the capabilities of human interpretation team. In this Human-in-the-Loop workflow (Figure 3), the AI acts as a drafting partner to reduce the cognitive load of the "zero-to-one" specification process. When new regulatory text is ingested, the HUMBR pipeline processes the content and generates a Candidate Interpretation (the suggestion). This suggestion is presented to human subject matter experts (SMEs) alongside the source text. The expert’s role shifts from drafting specifications from scratch to reviewing, validating, and refining the AI’s suggestions. This workflow ensures that while the AI accelerates the extraction of complex logic, the final sign-off authority remains strictly with human experts. Consequently, the evaluation presented in this section assesses the quality of these AI Suggestions against interpretations historically drafted entirely by humans. The goal is to determine if the AI
Reducing Hallucination in Enterprise AI Workflows via Hybrid Utility Minimum Bayes Risk (HUMBR)
suggestions achieve a level of fidelity that allows them to serve as a reliable starting point for high-stakes compliance engineering.
Regulatory Text
HUMBR Pipeline
AI Suggestion (Draft)
Human Expert Review & Revision
Final Spec
Figure 3: Human-in-the-Loop Workflow. The AI generates a suggestion which serves as a draft for the Human Expert.
5.2
Pipeline Setup
We constructed a heterogeneous ensemble of four common Large Language Models, comprising both proprietary and open-weights models ( Gemini-2.5-Pro, Gemini-2.5-Flash, Llama-3.3-70B, Llama-4-Maverick-17B-128E) with distinct training corpus. As shown in Section 3.3.3, to capture diverse reasoning paths, each model generated responses at four distinct temperature settings {0, 0.25, 0.5, 0.75}, resulting in a candidate pool of 𝑁 = 16 outputs per query. While we explored various ensemble sizes and utility weights during initial development, we found that the system performance remains robust across a reasonable range of parameters. For the consensus mechanism, we configured the Hybrid Utility function with a weighting of 𝛼 = 0.65 (favoring semantic consistency) and a strict consensus threshold of 𝜏 = 0.8 to prioritize high precision. The semantic embedding similarity was computed using opensource DRAMA embedding model, selected for its performance in retrieving nuanced legal concepts. Note that due to the nature of the proprietary regulatory framework, the specific prompts and raw legal texts used in this deployment cannot be publicly disclosed. 5.2.1 Evaluation. To validate whether these offline theoretical gains translate to tangible improvements in a high-stakes enterprise environment, we conducted a blind evaluation comparing three approaches: (1) Human: Interpretations drafted by internal and external legal counsel without AI suggestions. (2) Universal Self-Consistency (USC): A strong prompt-based baseline where an LLM aggregates the candidate pool to identify and refine the semantic consensus. (3) MBR Ensemble (Ours): The proposed centroid-based selection using the configurations described above. The evaluation set consisted of independent regulatory chunks. Given the prohibitive cost of legal expertise and the low-volume, high-value nature of regulatory workflows, we deliberately prioritized annotation quality over scale. Unlike crowd-sourced benchmarks, we engaged two legal experts to conduct a rigorous doubleblind review. To ensure an indisputable "Golden Standard," we exclusively utilized the subset where both experts reached full consensus after cross-verification. This strict filtering ensures that while the sample size is modest, the ground truth represents the highest tier of expert judgment essential for near-zero tolerance use case.
5.3
KDD ’26, August 9–13, 2026, Jeju, South Korea
Results: Preference and Win-Rates
As summarized in Table 3, the HUMBR ensemble demonstrated superior performance. Despite the conservative sample size inherent to expert-verified legal datasets, the performance gap between MBR and human baselines was sufficiently large to yield statistical significance (𝑝 < 0.05) based on 95% Wilson confidence intervals. This statistical robustness confirms that the observed improvements are a genuine signal of architectural superiority rather than an artifact of variance. Table 3: Pairwise Win-Rates in Blind Expert Evaluation. HUMBR outperforms with statistical significance Human baselines and heuristic ensembles. Statistical significance was verified using 95% Wilson confidence intervals. Comparison
Win
Loss
Win Rate (%)
HUMBR vs. Human Expert* HUMBR vs. Universal Self-Consistency* Universal Self-Consistency vs. Human
30 17 11
7 2 6
81.0% 89.5% 64.7%
* Indicates statistical significance at the 95% level based on Wilson confidence intervals.
Notably, the HUMBR achieved an 81% win rate against human experts working without such AI suggestions. Qualitative feedback indicated a perception that while human experts occasionally missed key sections of the law (a “recall” failure), the HUMBR system provided comprehensive coverage while maintaining conciseness.
5.4
Hallucination and Failure Mode Analysis
One requirement for this pipeline is the minimization of extrinsic hallucinations—the introduction of information not present in the source text. We categorized errors into rigorous failure modes based on a standardized rubric. Given the resource constraints of highexpert evaluation, we prioritized a deep-dive comparison solely between the Human and the top-performing candidate HUMBR. Figure 4 visualizes the complete distribution of outcomes for the two contending groups. HUMBR achieved the highest “No Failure Detected” rate (44.5%), nearly double that of human annotators (28.0%). Detailed analysis of the failure modes reveals a trade-off in zerotolerance domains: (1) Reduction of Critical Hallucinations: While LLMs are prone to generating text, HUMBR successfully filters out a notable type of error: Direct Contradictions. HUMBR reduced the rate of contradicting the source text from human level 4.7% to 2.5%. (2) Completeness and Definition Recall: As shown by the gray segments in Figure 4, human drafts were flagged as missing key sections of the source text in 28% of cases. In contrast, HUMBR reduced this specific error mode to negligible levels (0.8%). This disparity reflects a difference in drafting protocols. While human experts prioritize structural efficiency by defining terms globally to avoid redundancy, the HUMBR system generates self-contained explanations. In a blind evaluation of isolated text chunks, this self-contained
KDD ’26, August 9–13, 2026, Jeju, South Korea
HUMBR
Fang and Mola, et al.
44.5%
28%
Human Expert
0
28%
20 40 60 80 Distribution of Outcomes (%) Success Critical Halluc.
100
Recall Error Citation Issue
Figure 4: Outcome Distribution Analysis. The stacked bars represent the full breakdown of 100% of samples. HUMBR (Top) greatly expands the “Success” rate (Teal) and reduces the “Recall Errors” (Gray).
nature ensures comprehensive coverage that requires no external cross-referencing1 . (3) The Citation Trade-off: The increase in the “Citation/Style Issue” category (Orange) for HUMBR is primarily driven by “Uncited References” (25.2% for HUMBR vs 12.4% for Human). Qualitative analysis shows this often involves the model calculating specific numbers implied but not explicitly stated in the text. While technically an extrinsic hallucination, this is a precision error rather than a logic error, and is sounder than the omission errors observed in human baselines. Table 4 details the specific breakdown of these failure modes. Table 4: Detailed Failure Mode Breakdown. HUMBR dominates in quality and completeness.
Category
Human Expert
HUMBR
Success (No Failure)
28.0%
44.5%
Critical Errors Misses Key Sections Contradicts Source
28.0% 4.7%
0.8% 2.5%
Precision & Style Errors Uncited References Too Verbose / Long-winded Increased Complexity Other / Restates
12.4% 8.8% 9.8% 8.3%
25.2% 11.8% 4.2% 11.0%
Total
100.0%
100.0%
5.4.1 Cost-Benefit Analysis. While ensembling increases computational overhead (𝑁 = 16 calls), the operational cost remains viable for important workflows. Since the 𝑁 = 16 generations are independent by design, the architecture is highly parallelizable. Consequently, in terms of time, increasing 𝑁 does not increase end-to-end generation latency. The value of HUMBR is driven by efficiency. By providing a high-fidelity initial draft, the system helps transform the legal expert’s workflow from time-consuming de novo drafting to rapid verification, with limited editing based on professional experience. Preliminary metrics indicate that this "Draft-then-Verify" workflow reduces the time required for regulatory interpretation from hours to minutes. Consequently, expert throughput is expected to increase which should provide a balance to the computational cost involved. The marginal compute cost is negligible compared to the benefit of potential time saved for in-house counsel, effectively allowing legal teams to scale their expertise across a broader volume of compliance requirements without compromising quality.
6
Open Research Challenges and Future Work
Our results show the potential of the overall Minimum Bayes Risk approach, but the results and real world impact suggest several important directions and open challenges for the research community. Model diversity (which becomes a measurable parameter, 𝜌), should be optimized in deployment. By decreasing 𝜌 we can reduce the rate of growth of 𝐾 × 𝑀, against the Hallucination Tolerance, 𝜖. Our theoretical analysis highlights this is a driver of the cost/benefit tradeoff, while our empirical results provide a way to measure the engineering trade-off between efficiency (in terms of inference requests required) and effectiveness (ever tighter tolerance). More research is needed on techniques to increase diversity as measured by Intra-Model Correlation. Naturally using different language models with different training will tend to achieve this goal, but future work should also consider other potentially more efficient ways to reduce correlation, such as dynamic diversity maximization, prompt engineering, and different methods of output probability distribution.
7
Conclusion
In this paper, we presented a robust framework for reducing hallucinations in enterprise LLM workflows using Minimum Bayes Risk selection. By leveraging the statistical consensus of heterogeneous ensembles and a hybrid utility function, our method effectively filters out stochastic errors. We provided a theoretical guarantee for its optimality and demonstrated its superior performance on both public benchmarks and internal tasks. Our findings suggest that for high-stakes domains, HUMBR provides a scalable, reference-free path to trusted AI.
Acknowledgments 1 The blind evaluation presented discrete portions of interpretations in isolation. Conse-
quently, human adherence to standard drafting protocols (defining terms once globally rather than repeating them) was perceived by reviewers as missing content within the specific chunk. The AI, lacking global document awareness, repeated definitions inline, resulting in higher perceived completeness for the test format.
We extend our gratitude to Gabriel Forgues, Rajeev Rao, Derek Larson, Wendy Summer, Neel Reddy Pochareddy, Matt Sarmiento, Yanqing Peng, Shitong Zhu, Allison Zhang, Nathaniel Taylor, Pouyan Ghasemi, Xiaoning Yang, Meenakshi Tripathy, Ian Kaufman, Ann
Reducing Hallucination in Enterprise AI Workflows via Hybrid Utility Minimum Bayes Risk (HUMBR)
Lu, Shruthi Katakam, Miles Wu, Feiyue Wu, Zulka Gavlovski, Bryce Junkins, Raymond Shanniu Li, Michael Marcusa, Anuj Patwardhan, Lynn Richmond, Nick Manzoli, Rachel Villari, Chloe Lu, and Emily Van Deuren for their invaluable support.
References [1] Xinyun Chen, Renat Aksitov, Uri Alon, Jie Ren, Kefan Xiao, Pengcheng Yin, Sushant Prakash, Charles Sutton, Xuezhi Wang, and Denny Zhou. 2023. Universal Self-Consistency for Large Language Model Generation. arXiv:2311.17311 [cs.CL] https://arxiv.org/abs/2311.17311 [2] Yung-Sung Chuang, Yujia Xie, Hongyin Luo, Yoon Kim, James Glass, and Pengcheng He. 2024. DoLa: Decoding by Contrasting Layers Improves Factuality in Large Language Models. arXiv:2309.03883 [cs.CL] https://arxiv.org/ abs/2309.03883 [3] Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, Christopher Hesse, and John Schulman. 2021. Training Verifiers to Solve Math Word Problems. arXiv:2110.14168 [cs.LG] https://arxiv.org/abs/2110.14168 [4] Shehzaad Dhuliawala, Mojtaba Komeili, Jing Xu, Roberta Raileanu, Xian Li, Asli Celikyilmaz, and Jason Weston. 2023. Chain-of-Verification Reduces Hallucination in Large Language Models. arXiv:2309.11495 [cs.CL] https://arxiv.org/abs/ 2309.11495 [5] Yilun Du, Shuang Li, Antonio Torralba, Joshua B. Tenenbaum, and Igor Mordatch. 2023. Improving Factuality and Reasoning in Language Models through Multiagent Debate. arXiv:2305.14325 [cs.CL] https://arxiv.org/abs/2305.14325 [6] Bryan Eikema and Wilker Aziz. 2020. Is MAP decoding all you need? the inadequacy of the mode in neural machine translation. arXiv preprint arXiv:2005.10283 (2020). [7] Chenhao Fang, Derek Larson, Shitong Zhu, Sophie Zeng, Wendy Summer, Yanqing Peng, Yuriy Hulovatyy, Rajeev Rao, Gabriel Forgues, Arya Pudota, Alex Goncalves, and Hervé Robert. 2024. Ingest-And-Ground: Dispelling Hallucinations from Continually-Pretrained LLMs with RAG. arXiv:2410.02825 [cs.CL] https://arxiv.org/abs/2410.02825 [8] Chenhao Fang, Xiaohan Li, Zezhong Fan, Jianpeng Xu, Kaushiki Nag, Evren Korpeoglu, Sushant Kumar, and Kannan Achan. 2024. LLM-Ensemble: Optimal Large Language Model Ensemble Method for E-commerce Product Attribute Value Extraction. arXiv:2403.00863 [cs.IR] https://arxiv.org/abs/2403.00863 [9] Chenhao Fang, Yanqing Peng, Rajeev Rao, Matt Sarmiento, Wendy Summer, Arya Pudota, Alex Goncalves, Jordi Mola, and Hervé Robert. 2025. Privacy Artifact ConnecTor (PACT): Embedding Enterprise Artifacts for Compliance AI Agents. arXiv:2507.21142 [cs.CR] https://arxiv.org/abs/2507.21142 [10] Markus Freitag, Behrooz Ghorbani, and Patrick Fernandes. 2023. Epsilon Sampling Rocks: Investigating Sampling Strategies for Minimum Bayes Risk Decoding for Machine Translation. In Findings of the Association for Computational Linguistics: EMNLP 2023, Houda Bouamor, Juan Pino, and Kalika Bali (Eds.). Association for Computational Linguistics, Singapore, 9198–9209. doi:10.18653/v1/2023.findings-emnlp.617 [11] Vaibhava Goel. 2002. Minimum Bayes-risk automatic speech recognition. (2002). [12] Ziwei Ji, Nayeon Lee, Rita Frieske, Tiezheng Yu, Dan Su, Yan Xu, Etsuko Ishii, Ye Jin Bang, Andrea Madotto, and Pascale Fung. 2023. Survey of Hallucination in Natural Language Generation. Comput. Surveys 55, 12 (March 2023), 1–38. doi:10.1145/3571730 [13] Dongfu Jiang, Xiang Ren, and Bill Yuchen Lin. 2023. LLM-Blender: Ensembling Large Language Models with Pairwise Ranking and Generative Fusion. arXiv:2306.02561 [cs.CL] https://arxiv.org/abs/2306.02561 [14] Lorenz Kuhn, Yarin Gal, and Sebastian Farquhar. 2023. Semantic Uncertainty: Linguistic Invariances for Uncertainty Estimation in Natural Language Generation. arXiv:2302.09664 [cs.CL] https://arxiv.org/abs/2302.09664 [15] Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen tau Yih, Tim Rocktäschel, Sebastian Riedel, and Douwe Kiela. 2021. Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. arXiv:2005.11401 [cs.CL] https://arxiv.org/abs/ 2005.11401 [16] Kenneth Li, Oam Patel, Fernanda Viégas, Hanspeter Pfister, and Martin Wattenberg. 2024. Inference-Time Intervention: Eliciting Truthful Answers from a Language Model. arXiv:2306.03341 [cs.LG] https://arxiv.org/abs/2306.03341 [17] Tian Liang, Zhiwei He, Wenxiang Jiao, Xing Wang, Yan Wang, Rui Wang, Yujiu Yang, Shuming Shi, and Zhaopeng Tu. 2024. Encouraging Divergent Thinking in Large Language Models through Multi-Agent Debate. arXiv:2305.19118 [cs.CL] https://arxiv.org/abs/2305.19118 [18] Nelson F. Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang. 2023. Lost in the Middle: How Language Models Use Long Contexts. arXiv:2307.03172 [cs.CL] https://arxiv.org/abs/2307.03172 [19] Xueguang Ma, Xi Victoria Lin, Barlas Oguz, Jimmy Lin, Wen-tau Yih, and Xilun Chen. 2025. DRAMA: diverse augmentation from large language models to
KDD ’26, August 9–13, 2026, Jeju, South Korea
smaller dense retrievers. arXiv preprint arXiv:2502.18460 (2025). [20] Potsawee Manakul, Adian Liusie, and Mark J. F. Gales. 2023. SelfCheckGPT: ZeroResource Black-Box Hallucination Detection for Generative Large Language Models. arXiv:2303.08896 [cs.CL] https://arxiv.org/abs/2303.08896 [21] Bo Ni, Zheyuan Liu, Leyao Wang, Yongjia Lei, Yuying Zhao, Xueqi Cheng, Qingkai Zeng, Luna Dong, Yinglong Xia, Krishnaram Kenthapadi, Ryan Rossi, Franck Dernoncourt, Md Mehrab Tanjim, Nesreen Ahmed, Xiaorui Liu, Wenqi Fan, Erik Blasch, Yu Wang, Meng Jiang, and Tyler Derr. 2025. Towards Trustworthy Retrieval Augmented Generation for Large Language Models: A Survey. arXiv:2502.06872 [cs.CL] https://arxiv.org/abs/2502.06872 [22] Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul Christiano, Jan Leike, and Ryan Lowe. 2022. Training language models to follow instructions with human feedback. arXiv:2203.02155 [cs.CL] https://arxiv.org/abs/2203.02155 [23] Weijia Shi, Xiaochuang Han, Mike Lewis, Yulia Tsvetkov, Luke Zettlemoyer, and Scott Wen tau Yih. 2023. Trusting Your Evidence: Hallucinate Less with Contextaware Decoding. arXiv:2305.14739 [cs.CL] https://arxiv.org/abs/2305.14739 [24] Mirac Suzgun, Luke Melas-Kyriazi, and Dan Jurafsky. 2023. Follow the wisdom of the crowd: Effective text generation via minimum Bayes risk decoding. In Findings of the Association for Computational Linguistics: ACL 2023. 4265–4293. [25] Miles Turpin, Julian Michael, Ethan Perez, and Samuel R. Bowman. 2023. Language Models Don’t Always Say What They Think: Unfaithful Explanations in Chain-of-Thought Prompting. arXiv:2305.04388 [cs.CL] https://arxiv.org/abs/ 2305.04388 [26] Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. 2023. Self-Consistency Improves Chain of Thought Reasoning in Language Models. arXiv:2203.11171 [cs.CL] https://arxiv.org/abs/2203.11171 [27] Yue Zhang, Yafu Li, Leyang Cui, Deng Cai, Lemao Liu, Tingchen Fu, Xinting Huang, Enbo Zhao, Yu Zhang, Chen Xu, Yulong Chen, Longyue Wang, Anh Tuan Luu, Wei Bi, Freda Shi, and Shuming Shi. 2025. Siren’s Song in the AI Ocean: A Survey on Hallucination in Large Language Models. arXiv:2309.01219 [cs.CL] https://arxiv.org/abs/2309.01219 [28] Shitong Zhu, Chenhao Fang, Derek Larson, Neel Reddy Pochareddy, Rajeev Rao, Sophie Zeng, Yanqing Peng, Wendy Summer, Alex Goncalves, Arya Pudota, and Hervé Robert. 2025. Compliance Brain Assistant: Conversational Agentic AI for Assisting Compliance Tasks in Enterprise Environments. arXiv:2507.17289 [cs.AI] https://arxiv.org/abs/2507.17289
A
Proof of Theorem 1
Here we provide the detailed proof for MBR optimality mentioned in Section 3.1. ˆ 𝑦 ∗ ) be Theorem A.1 (Optimality of MBR Selection). Let U (𝑦, a utility function quantifying the quality of hypothesis 𝑦ˆ against the ground truth 𝑦 ∗ . If candidates are drawn from a distribution 𝑃𝜃 (𝑦|𝑥) that approximates the true posterior, then 𝑦ˆMBR maximizes the lower bound of the expected utility under the true distribution. Proof. Our objective is to find the hypothesis that maximizes the expected utility under the true posterior: ˆ 𝑦 ∗ )] 𝑦ˆ𝑜𝑝𝑡 = arg max E𝑦 ∗ ∼𝑃 (· |𝑥 ) [U (𝑦,
(14)
𝑦ˆ
Using the sample set C as a proxy for the true posterior (Monte Carlo approximation), the empirical MBR score is given by: 1 ∑︁ ˆ = ˆ 𝑐𝑖 ) 𝑆 (𝑦) U (𝑦, (15) 𝑁 𝑐 ∈C 𝑖
Consider the decomposition of the utility space. We assume correct answers {𝑦𝑐𝑜𝑟 } cluster in a high-density semantic region Ω𝑐𝑜𝑟 , while hallucinations {𝑦ℎ𝑎𝑙 } are scattered in low-density regions (the "sparse hallucination" assumption). (1) For a correct candidate 𝑦𝑐𝑜𝑟 ∈ Ω𝑐𝑜𝑟 , the expected score is dominated by consistency with other correct samples: E[𝑆 (𝑦𝑐𝑜𝑟 )] ≈ 𝑃 (𝑐𝑜𝑟 ) · Uℎ𝑖𝑔ℎ + 𝑃 (ℎ𝑎𝑙) · U𝑙𝑜𝑤
KDD ’26, August 9–13, 2026, Jeju, South Korea
Algorithm 1: Reference-Free MBR Selection with Consensus Threshold Input: Candidates C = {𝑐 1, . . . , 𝑐 𝑁 }; Weight 𝛼 ∈ [0, 1]; Consensus Threshold 𝜏 ∈ [0, 1] Output: Selected candidate 𝑐ˆ or Abstain 𝑁 ×𝑁 ← 0; Score vector 1 Init: Utility matrix 𝑈 ∈ R 𝑁 𝑆 ∈ R ← 0 // 1. Pre-compute Embeddings 2 𝐸 ← GetEmbeddings(C) ; // Matrix of size 𝑁 × 𝐷 // 2. Construct Pairwise Utility Matrix 3 for 𝑖 ← 1 to 𝑁 do 4 for 𝑗 ← 1 to 𝑁 do 5 if 𝑖 ≠ 𝑗 then 6 𝑢𝑠𝑒𝑚 ← CosSim(𝐸𝑖 , 𝐸 𝑗 ); 7 𝑢𝑙𝑒𝑥 ← ROUGE-L(𝑐𝑖 , 𝑐 𝑗 ); 8 𝑈𝑖 𝑗 ← 𝛼 · 𝑢𝑠𝑒𝑚 + (1 − 𝛼) · 𝑢𝑙𝑒𝑥 ; 9 end 10 end 11 end // 3. Compute Consensus Score 12 for 𝑖 ← 1 to 𝑁 do Í 13 𝑆𝑖 ← 𝑁 1−1 𝑗≠𝑖 𝑈𝑖 𝑗 ; // Avg distance to consensus 14 end // 4. Select Centroid with Consensus Gate ∗ 15 𝑘 ← argmax𝑘 ∈ {1,...,𝑁 } 𝑆𝑘 ; 16 if 𝑆𝑘 ∗ < 𝜏 then 17 return Abstain ; // Consensus too weak 18 end 19 return 𝑐ˆ = 𝑐 𝑘 ∗
(2) For a hallucinated candidate 𝑦ℎ𝑎𝑙 , consistency is low with both correct answers and other disjoint hallucinations: E[𝑆 (𝑦ℎ𝑎𝑙 )] ≈ 𝑃 (𝑐𝑜𝑟 ) · U𝑙𝑜𝑤 + 𝑃 (ℎ𝑎𝑙) · U𝑙𝑜𝑤 Since Uℎ𝑖𝑔ℎ > U𝑙𝑜𝑤 , it follows that E[𝑆 (𝑦𝑐𝑜𝑟 )] > E[𝑆 (𝑦ℎ𝑎𝑙 )]. ˆ is asymptotically Thus, maximizing the empirical consensus 𝑆 (𝑦) equivalent to identifying the mode of the true distribution 𝑃 (𝑦 ∗ |𝑥). □
B Hybrid Utility MBR Algorithm C Prompt Templates In this section, we provide the exact prompt templates used for the baseline comparisons and the evaluation metric.
Fang and Mola, et al.
Gold Reference Answer: [reference] AI Generated Answer: [answer] Criteria: 1. Truthfulness: Is the answer factually correct? Does it avoid common misconceptions? (0 or 1) 2. Informativeness: Does the answer provide relevant information(not just "I don't know")? (0 or 1) 3. Quality Score: A scalar score from 1-100 based on overall quality. Output a valid JSON object with the following keys: "truthful" (int), "informative" (int), "score" (int).
C.2
Prompt Template: Universal Self-Consistency (USC)
For the Universal Self-Consistency baseline, the following metaprompt was used to aggregate and select the most consistent answer from the candidate pool:
Prompt Template: Universal Self-Consistency (USC) I have generated the following responses to the question: [Prompted question or task] Response 1: [Response 1] Response 2: [Response 2] Response 3: [Response 3] ... Evaluate these responses. Select the most consistent response based on majority consensus. Start your answer with "The most consistent response is Response X" (without quotes).
C.3
Prompt Template: Contract NLI
Prompt Template: Contract NLI Analyze the following contract clause and hypothesis. Contract Clause: [premise]
C.1
Prompt Template: LLM-as-a-Judge
We utilized Claude-Opus-4.5 as an impartial judge to evaluate the Truthfulness and Informativeness of the generated answers. The evaluation logic is implemented as follows: Prompt Template: LLM-as-a-Judge You are an impartial expert judge. Evaluate the following AI-generated answer. Question: [question]
Hypothesis: [hypothesis] Based on the contract clause, determine if the hypothes is - "Entailment" (the clause supports the hypothesis) - "Contradiction" (the clause contradicts the hypothesis) - "Not mentioned" (the clause does not address the hypothesis) Answer with ONLY one word: Entailment, Contradiction, or NotMentioned.
Reducing Hallucination in Enterprise AI Workflows via Hybrid Utility Minimum Bayes Risk (HUMBR)
C.4
Prompt Template: Hearsay Detection
Prompt Template: Hearsay Detection Determine if the following statement constitutes hearsay under the Federal Rules of Evidence. Scenario: [text] Answer with ONLY: Yes or No.
C.5
Prompt Template: Rule QA
Prompt Template: Rule QA Answer the following legal question precisely. Question: [question] Provide a concise answer.
D
KDD ’26, August 9–13, 2026, Jeju, South Korea
Judge Reliability Validation
To ensure the reliability of our LLM-as-a-Judge evaluation presented in Section 4, we conducted the following two-stage validation procedures: Consistency Analysis. We re-evaluated a random 10% subset (82 samples) of TruthfulQA responses using the Claude-Opus-4.5 with identical prompts. The intra-judge agreement reached 93.3% on binary truthfulness decisions and a Pearson correlation of 𝑟 = 0.94 on the Quality Score, indicating high self-consistency. Alignment with Human Labels. We compared judge outputs against the official TruthfulQA human-annotated ground truth on 100 randomly sampled questions. The judge achieved 91.0% agreement with human annotators on the truthfulness classification, comparable to the inter-annotator agreement reported in the original benchmark.