arXiv:2606.09587v1 [cs.HC] 8 Jun 2026
Seeing the Hivemind: A Consensus-Aware Interaction Technique for Mitigating AI Homogenization Muhammad Haris Khan
Joel Wester
University of Copenhagen Denmark [email protected]
University of Copenhagen Denmark [email protected]
Figure 1: Overview of the Semantic Repulsion Technique (SRT) for mitigating AI homogenization in generative writing. (A) Standard LLM outputs for the same prompt often converge into similar high-probability responses. (B) SRT estimates this default response distribution by sampling the model K = 12 times, embedding the samples, computing the consensus centroid c, and extracting repeated phrases as negative concepts. (C) The Semantic Radar projects samples into a 2D landscape where the Yellow Zone marks common AI responses near c, not necessarily better responses; baseline outputs stay near c, while SRT outputs are generated farther away in lower-density regions. (D–E) A user-controlled repulsion strength 𝜆 drives the Repulsion Engine, combining contrastive decoding, phrase penalties, and fluency controls.
Abstract People are increasingly using AI for creative tasks such as writing. While adoption continues to grow, this form of use risks undermining individual creativity locally and reducing the heterogeneity of creative output at scale. In response, we introduce the Semantic Repulsion Technique (SRT) and evaluate it both computationally and through a study with 16 participants who regularly use AI for creative tasks. Our computational assessment reveals that SRT increases semantic diversity by 85–167% while reducing consensus phrases by 43–95% across task modes. In the user study, SRT outputs received higher usefulness (𝑝 = .019, 𝑊 = .208) and coherence
ratings ( 𝑝 = .006, 𝑊 = .260); 68.8% of participants were willing to use SRT-Strong for multiple tasks versus 18.8% for baselines. Originality and coherence ratings were positively correlated across all systems (𝜌 = +.40 to +.67), suggesting that divergence need not compromise readability. Taken together, these preliminary findings can inform the design of AI systems that aim to support everyday creativity without contributing to homogenization.
Khan et al.
CCS Concepts • Human-centered computing → HCI design and evaluation methods; User studies; Interactive systems and tools; • Computing methodologies → Natural language generation.
Keywords Human-agent interaction, human-AI collaboration, generative AI, large language models, AI-assisted creativity, user agency, consensus visualization, AI homogenization, creativity support tools, interactive AI systems
1
Introduction
When writers use AI for creative tasks, they face an invisible problem: their work begins to resemble others’ work. The rapid adoption of Large Language Models (LLMs) for writing assistance has lowered the barrier to generating fluent text, yet this democratization carries an unexpected cost. When different people use AI for the same creative task, their outputs converge toward remarkably similar responses [12]. This Artificial Hivemind effect emerges from how LLMs operate: generating text by predicting likely continuations from training data patterns. Alignment techniques amplify this convergence, as human evaluators favor typical outputs over unusual ones [28]. Across repeated samples for the same prompt, responses concentrate into a high-density region in semantic embedding space—the consensus. The result: AI assistance raises average quality while reducing variety and distinctiveness of ideas across users [1], [6]. This challenge raises pressing questions for HCI about human agency in relation to output convergence, and users’ awareness of how interactive systems shape divergence. Current methods obscure this convergence, leaving users unaware that their “original” ideas cluster with thousands of similar AI-assisted outputs. Default sampling keeps generation near high-probability modes, polished responses increase user acceptance [5], and acceptance reinforces convergence. Even exploration-oriented tools like Luminate [25] and Reverger [14] explore within the model’s conceptual map—if the model defaults to conventional ideas, exploring variations yields only refinements of familiar themes. Prior divergence strategies such as few-shot prompting or per-user fine-tuning shift the consensus without making it perceptible, and fine-tuning remains computationally impractical for everyday users. Users seeking creative distinction currently have no way to see or navigate away from algorithmic consensus. We ask: How do users perceive and evaluate AI-generated text that deliberately diverges from algorithmic consensus, and how does making consensus visible shape their understanding of AI homogenization? This motivates the Semantic Repulsion Technique (SRT), a research probe that operationalizes consensus-avoidance as a first-class interaction primitive. A critical framing distinction underlies our approach: SRT does not claim to generate creativity. Creativity requires human intent, personal meaning-making, and directed novelty that no repulsion parameter can encode. What SRT provides is divergence scaffolding—it makes the model’s default consensus region perceptible as a spatial object, then gives users controllable mechanisms to generate away from it. Whether users convert that divergence into genuinely creative work depends on
their own agency, judgment, and editorial effort. This distinction between consensus-avoidance and creativity proper is one we treat as empirically open and address directly in our evaluation. Rather than prompting models to “be creative,” SRT makes consensus visible and provides controls to navigate beyond it. For a prompt like “give me a sci-fi story premise,” while typical outputs cluster around familiar tropes, SRT visualizes the consensus zone as a “Yellow Zone” to support steering toward less probable but coherent alternatives. The system comprises three components: the Hivemind Detector samples multiple responses to identify default patterns; the Semantic Radar visualizes consensus as a “Yellow Zone” on a two-dimensional map; and the Repulsion Engine generates text that increases distance from this zone while preserving coherence. Users control divergence strength through a slider (𝜆 ∈ [0, 1.5]), making hidden algorithmic bias visible and navigable. We validate SRT through three complementary assessments designed to establish both what the system does and precisely how each of its components contributes. A controlled computational comparison across 1,500 generated outputs demonstrates that SRT increases semantic divergence by 85–167% while reducing consensus phrases by 43–95% relative to baseline sampling. A mechanism ablation study across 1,200 generations then isolates the causal contribution of each component: contrastive decoding drives divergence (+0.43 originality), phrase penalties provide surgical cliché suppression (−0.38 cliché frequency) at near-zero coherence cost, and fluency controls are the essential stabilizer that prevents the contrastive signal from producing incoherent output—reducing perplexity from ∼54,000 to ∼23. A further 𝜆-sweep study across 750 generations characterizes the divergence–relevance–coherence trade-off across the full operating range, establishing 𝜆 ≈ 1.2–1.8 as the practical ceiling beyond which originality plateaus while prompt relevance continues to decline. These computational results motivate and contextualize a user study with 16 participants that examines how users perceive, understand, and leverage consensus visualization in practice. Our contributions are threefold. First, we introduce Semantic Repulsion as an interaction technique for generative AI systems and demonstrate its implementation through SRT—to our knowledge the first system to operationalize consensus-avoidance as a first-class interaction objective. Second, through a three-part computational evaluation including mechanism ablation and operating-range analysis, we show that SRT produces measurably more divergent outputs while maintaining coherence, and characterize which mechanisms drive which outcomes. Third, through user evaluation, we examine how users perceive and respond to consensus visualization, and demonstrate that SRT reduces semantic homogenization in outputs compared to standard AI assistance—while acknowledging that divergence and creativity, though related, are empirically distinct constructs that our metrics address separately.
2 Related Work 2.1 Homogeneity in AI-Assisted Creativity When different users prompt LLMs for the same creative task, outputs converge toward similar responses. Jiang et al. [12] found both intra-model repetition and inter-model convergence across GPT4 [21], Claude 3 [2], and Llama 3 [20]. Zhang et al. [28] explain
Seeing the Hivemind: A Consensus-Aware Interaction Technique for Mitigating AI Homogenization
this theoretically: alignment training favors familiar responses, truncating distributional tails where novelty resides. The human cost is documented: AI assistance improves average quality but reduces semantic variance across groups [1], enhancing individual output while diminishing collective diversity [6]. Work on pluralistic alignment [24] inspires our approach of treating consensus as navigable rather than inevitable.
2.2
Novelty, Distance, and Divergence in Creativity Research
The relationship between semantic distance and creative output is well established in both cognitive science and computational creativity. Gupta et al. [10] demonstrated through the Remote Associates Test that creative solutions require active avoidance of highfrequency responses: individuals biased toward statistically likely answers systematically underperform, establishing that divergence from consensus is a prerequisite for creative problem-solving. Fu et al. [7] showed in engineering design that analogical distance follows a sweet-spot pattern—analogies too near the problem restrict novelty, while those too far become ineffective as inspiration; SRT’s tunable 𝜆 is directly motivated by this trade-off. Grace and Maher [9] formalize related intuitions computationally: their expectation-based novelty model defines an artefact as novel to the degree it violates a learned prior over likely outputs, a framing that closely parallels SRT’s operationalization of originality as distance from the consensus centroid. Kim and Maher [13] further show that the conceptual distance of AI-generated inspirations shapes the human ideation process itself, suggesting that where AI output sits in semantic space matters for what users produce next—motivating our approach of making that location visible and navigable.
Make consensus visible. Users cannot avoid consensus if they cannot see it. Current interfaces hide the model’s default tendencies, leaving users unaware that their “original” ideas may cluster with thousands of similar outputs. Our first goal is to make algorithmic consensus perceptible as a concrete object users can observe and reason about. Enable deliberate divergence with transparent control. Visualization alone is insufficient—users need mechanisms to act on what they see. Drawing on Rafner et al.’s emphasis on process transparency [22], we expose both the consensus landscape and the parameters that govern divergence strength, allowing users to make informed choices about how much novelty to pursue. Maintain coherence while diverging. Increasing distance from consensus risks generating incoherent or low-quality outputs. We balance novelty with readability through fluency constraints and targeted penalties, ensuring that divergent outputs remain useful rather than merely different.
3.2
Model Architecture
SRT employs a dual-model contrastive decoding architecture using two instruction-tuned causal language models: Qwen2.5-7BInstruct (strong model) and Qwen2.5-1.5B-Instruct (weak model) [27]. Both models support optional 4-bit NF4 quantization via bitsandbytes when CUDA is available. Since the two models use different tokenizers with vocabulary sizes |𝑉𝑠 | and |𝑉𝑤 |, a pre-computed vocabulary alignment mapping m ∈ Z |𝑉𝑠 | matches token strings between models. For each strong-model token index 𝑖, the mapping is 𝑚𝑖 = 𝑗 if the token string matches weak index 𝑗, and 𝑚𝑖 = −1 otherwise. Typical vocabulary overlap is 70–80%. Semantic similarity throughout the system is computed via all-MiniLM-L6-v2, which maps variable-length text to 3842.3 Creativity Support Tools and User Agency dimensional 𝐿2 -normalized embeddings, with cosine distance as the Building on Shneiderman [23], tools like Luminate [25] and Reverger [14] divergence metric: help users explore AI-generated possibilities through structured vari𝑑 (u, v) = 1 − u⊤ v (1) ation. However, these systems assume users want to explore within where ∥u∥ = ∥v∥ = 1. the model’s possibility space rather than escape it—they provide no way to see or navigate away from algorithmic consensus. Chen et al. [5] show polished AI outputs make rejection difficult, while Rafner et al. [22] emphasize preserving agency through process control. Our work differs fundamentally: rather than exploring within algorithmic boundaries, we make consensus visible and provide controls to navigate beyond it. To our knowledge, no existing system operationalizes consensus-avoidance as a first-class interaction objective.
3
System Design
We designed the Semantic Repulsion Technique (SRT) to address a fundamental challenge in human-AI co-creation: how can users recognize and navigate away from algorithmic consensus when seeking creative distinction? Our design translates an invisible statistical property—the model’s default response distribution—into visible, manipulable interface elements that support user awareness and control.
3.1
Design Goals
We designed SRT around three core principles:
3.3
User Workflow
SRT transforms single-shot generation into a multi-stage interaction: (1) Users enter prompts and select a task mode (Creative, Technical, or Brainstorm). (2) The Hivemind Detector generates 12 samples to estimate the consensus distribution. (3) The Semantic Radar visualizes consensus as a “Yellow Zone” on a 2D map. (4) Users adjust a repulsion slider (𝜆) and generate baseline versus repulsed outputs side-by-side, each annotated with its cosine distance from the consensus centroid. (5) Users compare results and iterate with different settings.
3.4
Hivemind Detector
Given a prompt 𝑝 and task mode M, the Hivemind Detector generates 𝐾 = 12 samples using nucleus sampling [11] (temperature 𝜏temp , top-𝑝 threshold 𝑝 nucleus , maximum 128 tokens). We chose 𝐾 = 12 as it provides sufficient samples for UMAP [19] projection (which requires 𝐾 ′ ≥ 6 unique samples) while remaining computationally tractable. After deduplication—removing exact string duplicates— 𝐾 ′ ≤ 𝐾 unique samples remain.
Khan et al.
Centroid and Modal Sample. Let {𝑠 1, . . . , 𝑠𝐾 ′ } denote unique samples and {e1, . . . , e𝐾 ′ } their 𝐿2 -normalized embeddings. The consensus centroid is: Í𝐾 ′ e𝑖 c = Í𝑖=1′ (2) 𝐾 𝑖=1 e𝑖
Brainstorm Mode uses a curated list of marketing frames (e.g., “revolutionary solution,” “game changer,” “cutting edge,” “think outside the box”) as a primary source, selected to target the overrepresentation of promotional framing in LLM brainstorming outputs [12]. If fewer than 12 concepts are found, the Creative mode YAKE extraction pipeline fills the remainder.
The modal sample minimizes cosine distance to c: 𝑖 ∗ = arg min 1 − e𝑖⊤ c
3.6 (3)
𝑖
Sample 𝑠𝑖 ∗ represents “what the model most wants to say.” The detector also extracts negative concepts—specific phrases characterizing consensus patterns.
3.5
Negative Concept Extraction
Negative concepts are consensus-characterizing phrases that the Repulsion Engine penalizes during generation. Extraction differs by task mode to reflect a fundamental difference in how consensus manifests across modes. In Creative mode, consensus takes the form of prompt-specific narrative clichés that vary with the topic (e.g., “AI uprising” for science fiction, “dystopian society” for speculative fiction), so phrases must be extracted dynamically from the sampled outputs. In Technical and Brainstorming modes, consensus manifests instead as prompt-invariant surface-form boilerplate—stylistic hedging and marketing formulae that appear regardless of the specific task content—making a curated list both sufficient and more reliable than dynamic extraction. This design choice is validated empirically: our ablation study (Section 4.2) shows that phrase penalties reduce consensus phrase frequency by 0.38 with near-zero coherence cost (ΔPerplexity = −89 relative to a baseline of 4.5), confirming that the mode-specific extraction strategy successfully targets surface-form consensus without disrupting fluency. Creative Mode uses YAKE [4] keyword extraction with maximum 𝑛-gram length 𝑛 max = 4 and candidate pool size 𝑘 top = 120, applied to concatenated sample text. Candidates are filtered through five stages: (1) Phrase-only filtering requires ≥ 2 alphabetic words per phrase. (2) Generic content filtering removes phrases where half or more content words belong to a predefined generic set (e.g., “time,” “people,” “things,” “way”) or to standard English stopwords. (3) Cross-sample consensus filtering retains only phrases appearing in ≥ 2 of the 𝐾 ′ samples, ensuring they represent genuine consensus patterns rather than one-off expressions. (4) Prompt exclusion discards phrases whose words are entirely contained in the user’s original prompt. (5) The remaining phrases are ranked by (frequency, −YAKE score) and the top 12 are selected. Technical Mode uses a curated list of stylistic boilerplate phrases (e.g., “let’s break down,” “let’s delve into,” “in summary,” “as you can see”) and returns only those observed in ≥ 1 sample, up to a maximum of 12 concepts. These phrases were selected based on documented patterns of LLM-generated hedging and transitional language that signal AI-produced text without contributing informational content [28].
Semantic Radar Visualization
The Semantic Radar transforms high-dimensional embedding space into an interactive 2D landscape that makes consensus visible as ′ a spatial object. We construct a matrix E ∈ R (𝐾 +1) ×384 containing sample embeddings and the centroid (if a user draft is provided, its embedding is appended). When sufficient unique samples exist (𝐾 ′ ≥ 6), we apply UMAP [18] with n_neighbors = min(10, 𝐾 ′ −1), min_dist = 0.15, cosine metric, and fixed random_state = 42. For smaller sets (𝐾 ′ < 6), we fall back to PCA with 2 components. Drawing on techniques from DataMap [8] and latent space visualization research [17], we repurpose dimensionality reduction for a human-centered goal: making algorithmic consensus visible as a spatial object users can navigate around. The signature “Yellow Zone” emerges from Gaussian kernel density estimation (KDE) over the projected sample positions. Let ′ X ∈ R𝐾 ×2 denote projected sample coordinates. The density estimate is: 𝐾′ 1 ∑︁ 𝑓ˆ(x) = ′ N x; X[𝑖], Σ (4) 𝐾 𝑖=1 where Σ is estimated via Scott’s rule. We evaluate 𝑓ˆ on a 120 × 120 grid spanning the data range ±0.8 margin, and render it as a semitransparent contour plot (opacity = 0.35), where darker regions indicate higher consensus density. Sample points are displayed as scatter markers with hover text showing full sample content, the centroid is marked distinctly showing the modal sample, and an optional “you” marker shows the user’s draft position if provided. The spatial metaphor (distance = semantic distance) relies on the underlying embedding model, where relative positions in the 2D projection preserve relative semantic distances from the original high-dimensional space. Key design principles include progressive disclosure (overview to detail), personal positioning (users overlay drafts to see their location relative to consensus), and transparent representation (the centroid shows the modal sample text).
3.7
Repulsion Engine
The Repulsion Engine provides a slider (“Baseline” to “Strong”) mapped to parameter 𝜆 ∈ [0, 1.5] governing semantic distance from consensus. The engine combines three complementary mechanisms: Contrastive Decoding. Inspired by [15] and [16], at each generation step 𝑡, we compute logits from both models and apply: ℓ ct(𝑡 ) = ℓ 𝑠(𝑡 ) − 𝜆 · ℓ 𝑤(𝑡 )
(5)
where ℓ 𝑠(𝑡 ) ∈ R |𝑉𝑠 | are strong model logits and ℓ 𝑤(𝑡 ) are weak model logits aligned to the strong vocabulary via mapping m: ( ℓ (𝑡 ) [𝑚𝑖 ] if 𝑚𝑖 ≥ 0 (𝑡 ) ℓ𝑤 [𝑖] = 𝑤,raw (6) 0 otherwise
Seeing the Hivemind: A Consensus-Aware Interaction Technique for Mitigating AI Homogenization
This amplifies the strong model’s distinctive capabilities while suppressing patterns both models share—effectively penalizing “consensus” predictions. Higher 𝜆 produces greater suppression of shared patterns and thus greater divergence. Phrase-Level Penalties. Building on unlikelihood training [26], the engine applies a direct logit penalty 𝛽 to all tokens belonging to negative concept phrases extracted by the Hivemind Detector: ℓct(𝑡 ) [𝑖] ← ℓct(𝑡 ) [𝑖] − 𝛽, ∀ 𝑖 ∈ Tneg
(7)
Ð where Tneg = ph∈ N tok(ph) is the union of all tokens appearing in negative concept phrases, and 𝛽 = 4.0 across all modes. Because the penalty operates at the token level, it suppresses not only exact consensus phrases but also their constituent words in other contexts— a deliberate trade-off favoring broader consensus avoidance over surgical precision. Fluency and Diversity Controls. Three mechanisms maintain output quality during divergence: (1) Fluency floor. Tokens whose strong-model probability falls below threshold 𝜏 are masked: ℓct(𝑡 ) [𝑖] ← −∞ if 𝑝𝑠(𝑡 ) (𝑖) < 𝜏
Table 1: Hyperparameters by task mode. Interface defaults are 𝜆 = 1.2 (Creative), 𝜆 = 0.6 (Technical), and 𝜆 = 1.5 (Brainstorm). For the controlled evaluation, 𝜆 = 0.6 (SRT-Mild) and 𝜆 = 1.2 (SRT-Strong) were applied uniformly across all modes.
(8)
where 𝑝𝑠(𝑡 ) (𝑖) = softmax(ℓ 𝑠(𝑡 ) ) [𝑖], with 𝜏 = 0.003 for Creative, 0.010 for Technical, and 0.002 for Brainstorm. (2) Repetition penalty. For each previously generated token 𝑖 ∈ y<𝑡 : ( ℓ (𝑡 ) [𝑖]/𝜌 if ℓct(𝑡 ) [𝑖] > 0 (𝑡 ) ℓct [𝑖] ← ct(𝑡 ) (9) ℓct [𝑖] · 𝜌 otherwise with 𝜌 = 1.06 for Creative/Brainstorm and 𝜌 = 1.12 for Technical. (3) 𝑁 -gram blocking. We set ℓct(𝑡 ) [𝑖] ← −∞ for any token 𝑖 that would complete a repeated 𝑛-gram (𝑛 = 4 for Creative/Brainstorm, 𝑛 = 6 for Technical). Additionally, tokens containing non-ASCII alphabetic characters (excluding a mathematical symbol whitelist Λ = {𝜆, 𝛾, 𝜇, 𝜎, 𝜋, 𝜃, 𝛼, 𝛽, 𝛿, 𝜖, 𝜅, 𝜌, 𝜈, 𝜏 }) are blocked throughout generation. After temperature scaling and top-𝑝 filtering, the final token is sampled from the resulting distribution. If the softmax produces NaN or zero sum, the engine falls back to strong-model-only sampling with the same constraints. In Creative mode, an optional polish pass uses greedy decoding (max 240 tokens) to fix spelling and grammar without reintroducing penalized phrases. Practical Interpretation of 𝜆. The slider provides predictable control: higher values consistently produce greater semantic distance from consensus. At 𝜆 = 0 (baseline), the system generates using the strong model alone. At 𝜆 = 0.6 (SRT-Mild), the weak model’s contribution is moderately subtracted, producing stylistic variation while staying relatively close to default patterns. At 𝜆 = 1.2 (SRT-Strong), the subtraction is substantial, yielding outputs with markedly different framing, vocabulary, and structure. Users see both baseline and repulsed outputs side-by-side, each annotated with its cosine distance from the centroid, enabling direct comparison of the divergence trade-off. Algorithm 1 summarizes the core Semantic Repulsion Technique (SRT) pipeline. Given a user prompt, SRT first samples the model’s default response distribution to estimate the consensus region, extracts consensus-characterizing phrases, and then performs repulsionbased contrastive decoding to generate an output that moves away from common model patterns while preserving fluency.
3.8
Parameter
Cre.
Tech.
Brain.
𝜆 (repulsion) 𝜏 (fluency floor) 𝛽 (phrase penalty) 𝜌 (rep. penalty) 𝑛 (ngram block) 𝐿gen (tokens) Baseline dec. Polish pass
1.2 0.003 4.0 1.06 4 320 Samp. Opt.
0.6 0.010 4.0 1.12 6 220 Greedy No
1.5 0.002 4.0 1.06 4 260 Samp. No
Task Modes
Mode-specific defaults reflect varying requirements for accuracy versus exploration. Table 1 summarizes the full configuration. Creative mode (𝜆 = 1.2) targets narrative clichés using YAKEextracted consensus phrases. Technical mode (𝜆 = 0.6) uses a higher fluency floor (𝜏 = 0.010) and stricter repetition constraints (𝜌 = 1.12, 𝑛 = 6) to preserve factual accuracy while targeting stylistic boilerplate. Brainstorm mode (𝜆 = 1.5) targets marketing frames and allows the most aggressive exploration with the lowest fluency floor (𝜏 = 0.002).
4
System Evaluation
We validate SRT through three complementary computational assessments: (1) a controlled comparison establishing that SRT produces measurably more divergent outputs than baseline sampling strategies; (2) a mechanism ablation isolating the causal contribution of each SRT component; and (3) an operating-range analysis characterizing the divergence–relevance–coherence trade-off across the full 𝜆 range. Together these address the core technical claims: that SRT diverges from consensus, that each mechanism serves a distinct and necessary role, and that the system remains prompt-relevant within its intended operating range.
4.1
Computational Assessment
Setup. We conducted a controlled comparison across 30 prompts (10 per task mode), comparing five systems: Baseline-Pure (nucleus sampling [11], temp = 1.0, top-𝑝 = 0.9), Baseline-HighTemp (temp = 1.5), Baseline-Beam (num_beams = 5), SRT-Mild (𝜆 = 0.6), and SRT-Strong (𝜆 = 1.2). All systems used Qwen2.5-7BInstruct; SRT variants additionally employed contrastive decoding with Qwen2.5-1.5B-Instruct as the weak model. Values of 𝜆 were applied uniformly across all modes to isolate the effect of contrastive repulsion. For each prompt, the Hivemind Detector established consensus by generating 12 samples, computing 384-dimensional embeddings via all-MiniLM-L6-v2, calculating the consensus centroid, and extracting negative concepts via YAKE [4]. We then generated 10 outputs per system per prompt (1,500 total). Metrics. We measured three complementary aspects. Originality (distance from consensus centroid): 𝐷 mode (𝑔) = 1 − e𝑔⊤ c
(10)
Khan et al.
Higher values indicate greater divergence from defaults. Diversity (intra-system): mean pairwise cosine distance between all outputs from the same system for the same prompt. Consensus Phrase Frequency: count of negative concepts appearing in each output; lower values indicate greater consensus suppression. We note that embedding-based originality measures semantic distance from the consensus but cannot distinguish diversity of expression from diversity of ideas; this limitation is acknowledged in Section 5. Results. Results strongly support SRT’s divergence capabilities across all three task modes (Figure 3). Originality. SRT-Strong achieved substantially higher divergence than all baselines across every mode. In Creative mode, SRT-Strong outputs were nearly twice as distant from the centroid as BaselinePure (0.50 vs. 0.27)—an 85% improvement. Technical mode showed the largest relative gain (+167%), and Brainstorm mode showed +119% improvement. SRT-Mild showed intermediate performance in all modes, demonstrating controllable divergence via 𝜆. BaselineBeam exhibited the lowest divergence across all modes, confirming that probability maximisation produces highly consensual outputs. Baseline-HighTemp showed high variance without consistently higher originality, suggesting unfocused divergence. Diversity. SRT maintained or improved intra-system diversity compared to sampling baselines, demonstrating broader semantic exploration rather than convergence to a new narrow peak. In Creative mode, SRT-Strong achieved diversity of 0.49 compared to 0.40 for Baseline-Pure while simultaneously achieving higher originality— evidence against simple mode substitution. Baseline-Beam showed zero diversity by design. Consensus Phrase Frequency. SRT-Strong achieved near-zero counts in Creative mode (0.08)—a 95% reduction versus BaselinePure (1.70) and 98% versus Baseline-Beam (3.30). Brainstorm mode showed 93% reduction. Technical mode exhibited higher absolute counts due to boilerplate detection methodology, but SRT-Strong still reduced consensus phrases by 43% versus Baseline-Pure. Mode-Specific Effectiveness. Creative and Brainstorm modes, which tolerate greater semantic exploration, showed larger originality gains. Technical mode, configured with a higher fluency floor (𝜏 = 0.010) and stricter repetition constraints to preserve factual accuracy, demonstrated meaningful divergence while maintaining coherence—evidence that the system successfully navigates the novelty–accuracy trade-off.
4.2
Mechanism Ablation
The computational assessment establishes that SRT diverges from consensus, but does not indicate which component drives which outcome. We address this through a 23 factorial ablation across 1,200 generated outputs (30 prompts × 8 conditions × 5 samples), toggling each of the three SRT mechanisms—Contrastive Decoding (CD), Phrase Penalties (PP), and Fluency Controls (FL)—independently. In addition to the three original metrics, we report two new metrics that directly concerns with prompt-relevance and objective coherence: Relevance, the cosine similarity between the prompt embedding and the output embedding (higher = more on-topic); and Coherence, the negative mean token log-likelihood under the strong model (higher = more fluent), also expressed as Perplexity (exp(−NLL); lower = more fluent).
Figure 2: The Semantic Radar’s Yellow Zone visualization. Gaussian KDE renders consensus density as a heatmap over UMAP-projected sample embeddings; darker regions indicate higher response clustering. The centroid (marked) represents the modal sample. Users can overlay their own drafts to see their position relative to consensus.
Results. Table 2 reports the main-effect delta for each mechanism— the mean difference in each metric when the mechanism is ON versus OFF, averaged across all conditions in which the other two mechanisms vary. Each mechanism serves a distinct and necessary role (Figure 4). Contrastive Decoding is the primary driver of divergence (ΔOriginality = +0.43) and the dominant source of cliché suppression (ΔCliché = −0.79). Used in isolation, however, CD produces catastrophically incoherent output—perplexity rises from a baseline of 4.5 to ∼54,000 (C1_CD only condition)—because the contrastive subtraction suppresses likely tokens without any floor on impossible ones. This confirms that CD is a powerful but unsafe mechanism when deployed alone. Phrase Penalties provide surgical cliché suppression (ΔCliché = −0.38) with near-zero impact on all other metrics (ΔPerplexity = −89, negligible relative to the baseline of 4.5; ΔOriginality = +0.01; ΔRelevance = −0.007). This cleanly justifies the mode-specific extraction design: phrase penalties target surface-form consensus without interfering with the embedding-space repulsion signal. Fluency Controls are the essential stabilizer: they reduce perplexity by 53,907 points (ΔCoherence = +4.16), recovering from the CD-induced incoherence while accepting a modest originality cost (ΔOriginality = −0.29). The positive relevance effect (ΔRelevance = +0.19) shows that fluency controls also restore on-topic quality. The full SRT configuration (C7) balances these contributions: in Creative mode it achieves originality 0.48 versus baseline 0.29 (+65%), reduces cliché frequency from 1.38 to 0.04 (−97%), and maintains perplexity at 71—substantially higher than the baseline of 5.9 but five orders of magnitude lower than CD alone. All Wilcoxon signed-rank comparisons of C7 versus C0 (baseline) are 𝑝 < 0.001 across all metrics and modes, confirming that the observed effects are not attributable to sampling variance.
4.3
Operating Range Analysis
A key open question is how far repulsion strength 𝜆 can be increased before outputs lose prompt relevance. We address this through a 𝜆sweep study across nine values 𝜆 ∈ {0.0, 0.6, 1.2, 1.8, 2.4} with phrase penalties and fluency controls held fixed, generating 750 outputs (30
Seeing the Hivemind: A Consensus-Aware Interaction Technique for Mitigating AI Homogenization
Figure 3: Computational evaluation results across task modes. (a) Originality: distance from consensus centroid. (b) Diversity: intrasystem variance. (c) Consensus Phrase Frequency: count of consensus phrases. SRT-Strong (green) achieves 85%, 167%, and 119% originality improvements over Baseline-Pure (gray) in Creative, Technical, and Brainstorm modes. SRT-Mild (blue) shows intermediate performance, demonstrating controllable divergence via 𝜆. Error bars show standard deviation (10 samples per condition). Table 2: Mechanism ablation: main-effect deltas (ON − OFF), averaged across all conditions and modes (𝑁 = 150 per mechanism state). ΔCoherence is the change in negative NLL (higher = more fluent). Mechanism
ΔOrig.
ΔDiv.
ΔCliché
ΔRel.
ΔCoh.
Contrastive Decoding Phrase Penalties Fluency Controls
+0.434 +0.012 −0.287
+0.149 +0.002 −0.028
−0.785 −0.378 +0.058
−0.283 −0.007 +0.191
−6.03 −0.02 +4.16
Table 3: 𝜆-sweep summary averaged across all modes (𝑁 = 150 per 𝜆, mean values shown). 𝜆
Orig.
Rel.
Cliché
Coh.
PPL
0.0 0.6 1.2 1.8 2.4
0.184 0.253 0.325 0.365 0.370
0.559 0.510 0.481 0.461 0.458
0.513 0.113 0.073 0.047 0.020
−1.37 −2.22 −3.28 −3.80 −4.06
4.5 12.5 42.8 72.2 92.2
Figure 4: Mechanism ablation main-effect deltas. Each cluster shows the effect of enabling one mechanism (ON − OFF), averaged across the other mechanisms’ states. Coherence column is divided by 10 for display. Contrastive Decoding drives divergence but is unsafe alone (PPL → 54,000); Phrase Penalties provide surgical cliché suppression at near-zero coherence cost; Fluency Controls are the essential stabilizer.
Originality increases monotonically from 0.18 at 𝜆 = 0 to 0.36 at 𝜆 = 1.8 (+99%), then plateaus: the gain from 𝜆 = 1.8 to 𝜆 = 2.4 is only +0.005 (+1.4%). Relevance declines gradually, falling from 0.56 at baseline to 0.46 at 𝜆 = 2.4—a maximum drop of only 18.1%, indicating that outputs remain substantially on-topic across the full range tested. Computational coherence degrades more sharply (perplexity 4.5 → 92.2), though the user study shows that perceived coherence is higher for SRT-Strong than for the baseline (𝑀 = 3.92 vs. 3.48, 𝑝 = .006), suggesting that self-perplexity overestimates the readability cost of divergence for human readers. The diminishing returns in originality beyond 𝜆 = 1.8, combined with continued declines in relevance and coherence, establish 𝜆 ≈ 1.2–1.8 as the practical operating range. The interface defaults of 𝜆 = 1.2 (Creative), 𝜆 = 0.6 (Technical), and 𝜆 = 1.5 (Brainstorm) all fall within this empirically validated zone. Technical mode shows notably greater robustness: relevance drops only 8% from baseline even at 𝜆 = 2.4 (0.65 vs. 0.71), validating its lower default configuration. All Wilcoxon tests comparing each 𝜆 > 0 to 𝜆 = 0 are 𝑝 < 0.0001 across all metrics, confirming that even the smallest tested step (𝜆 = 0.6) produces statistically reliable divergence.
prompts × 5 𝜆 values × 5 samples). The same five metrics as the ablation study are reported. Results. Table 3 and Figure 5 show the divergence–relevance– coherence trade-off across the operating range.
To evaluate whether computational divergence translates into perceived utility and adoption intent, we conducted a within-subjects exploratory study with 𝑁 = 16 participants (ages 22–48; 4 female, 12 male) who regularly use AI for creative tasks (Table 4). We treat
4.4
User Study
Khan et al.
Figure 5: 𝜆-sweep: originality (green), relevance (blue), and coherence/−NLL (red, right axis) as a function of repulsion strength. Shaded region indicates the practical operating zone (𝜆 = 0.6–1.8) where originality gains are large and relevance remains above 82% of baseline. Error bars show SD (𝑁 = 150 per point). this as a preliminary investigation complementing the computational validation in Section 4; findings should be interpreted as suggestive rather than generalisable given the sample size. Procedure. Participants completed three tasks: (1) Creative Writing (rewriting a sci-fi story premise), (2) Technical Writing (explaining a topic of interest), and (3) Brainstorming (generating ideas). For each task, they reviewed four AI-generated responses—BaselinePure, Baseline-HighTemp, SRT-Mild, SRT-Strong—yielding 12 response evaluations per participant in total (3 tasks × 4 systems). Each system produced a single output per task to enable controlled within-subjects comparison across all four conditions simultaneously; iterative human–AI co-writing is an important direction we leave to future work. After reviewing all responses, participants rated each on 5-point Likert scales for originality, coherence, creativity, and usefulness, indicated willingness to use, and provided open feedback. Following all tasks, we presented the Yellow Zone visualization (Figure 2), explained consensus detection, and asked participants to select which system(s) they would most want to use for creative projects. System preferences were elicited after the Yellow Zone explanation, which may have primed participants toward consensus-aware systems; these results should therefore be interpreted alongside the blind Likert ratings. Analysis. We used non-parametric tests appropriate for ordinal Likert data and small samples. For each measure, we aggregated ratings across three tasks per participant-system combination, then
conducted Friedman tests to detect overall differences, followed by pairwise comparisons where warranted. Results. Figure 6 shows convergent evidence across multiple utility dimensions. Perceived Usefulness. SRT-Strong received highest usefulness ratings (𝑀 = 3.60), 16% higher than Baseline-Pure (𝑀 = 3.10). A Friedman test revealed significant differences, 𝜒 2 (3, 𝑁 = 16) = 9.99, 𝑝 = .019, Kendall’s 𝑊 = .208 (small-to-medium effect). While the sample size is modest, the effect size provides preliminary evidence of meaningful utility differences. Perceived Coherence. SRT-Strong also received significantly higher coherence ratings (𝑀 = 3.92, Mdn = 4.00) compared to BaselinePure (𝑀 = 3.48, Mdn = 3.50). A Friedman test revealed significant differences, 𝜒 2 (3, 𝑁 = 16) = 12.47, 𝑝 = .006, Kendall’s 𝑊 = .260 (medium effect). This finding is notable because it suggests that semantic divergence does not compromise—and may even enhance— perceived readability, a result consistent with the positive originality– coherence correlations observed below. Willingness to Use. Beyond ratings, behavioral intent showed a clear pattern favoring consensus-aware systems. At the participant level, 87.5% of participants (14/16) indicated willingness to use SRTStrong for at least one task, compared to 50.0% (8/16) for BaselinePure (Figure 6b). Under a stricter threshold—willingness for the majority of tasks (2+ of 3)—the gap widened further: 68.8% (11/16) for SRT-Strong versus only 18.8% (3/16) for Baseline-Pure, a 3.7× difference. A Cochran’s Q test for repeated binary measures showed
Seeing the Hivemind: A Consensus-Aware Interaction Technique for Mitigating AI Homogenization
Table 4: Participant demographics (𝑁 = 16).
as consensus clustering; 6/16 (38%) misinterpreted it as indicating optimal responses. Rather than treating this as a study failure, we Characteristic Summary read it as a design finding: users interpret visual density as quality Age rather than frequency. Future consensus visualizations should use Range 22–48 explicit labelling (e.g., “common” vs. “good”) and progressive disMean (SD) 28.1 (6.3) closure to prevent this conflation—a design implication independent 18–24 2 (12.5%) of whether SRT’s underlying divergence mechanism is effective. 25–30 11 (68.8%) Qualitative Results. We conducted thematic analysis [3] on 31–50 3 (18.8%) all open-ended responses, comprising per-response feedback (4 reGender sponses × 3 tasks) and 8 post-task reflection questions. Following Male 12 (75.0%) Braun and Clarke’s six-phase process, we familiarized ourselves Female 4 (25.0%) with the full dataset, generated initial codes, searched for candidate AI Tool Usage Frequency Daily 12 (75.0%) themes, reviewed themes against coded extracts, and refined definiWeekly 3 (18.8%) tions iteratively. Response depth varied substantially: 11 participants Rarely 1 (6.3%) provided detailed open-ended feedback, while 5 gave minimal or Creative Writing Frequency single-word responses on several items. Prevalence counts below reDaily 8 (50.0%) port the number of participants contributing at least one substantive Weekly 6 (37.5%) coded excerpt. Four themes were identified. Rarely 2 (12.5%) Before reporting themes, we note a key finding regarding visuAI Tools Used (multi-select) alization interpretability. When asked to interpret the Yellow Zone ChatGPT 12 (75.0%) before explanation, only 4/16 (25%) correctly identified it as consenClaude 6 (37.5%) sus clustering (e.g., P3: “a place where similar concepts clustered”). Gemini 5 (31.3%) Six participants (38%) misinterpreted it as indicating optimal reOther (DeepSeek, Copilot, etc.) 6 (37.5%) sponses (e.g., P9:“most accurate responses”). Theme 1: Coherence as Non-Negotiable Baseline (13/16). The a trend toward difference across systems, 𝑄 (3) = 6.91, 𝑝 = .075. most consistent pattern was that participants wanted distinctive outWhile not reaching conventional significance, the consistency of puts but treated coherence as a non-negotiable precondition. P13 preference across both thresholds suggests genuine adoption intent. captured this directly: “I need divergent writing but coherent—I System Preferences. When asked to select which system(s) they don’t want my essay to be same as others but I don’t want it to be would most want to use (participants could select multiple), 50% messy either.” P12 stated: “I would always choose coherence over chose SRT-Strong—the single most preferred option. Baseline-HighTemp creativity.” When outputs lacked readability, participants rejected and SRT-Mild were each selected by 38%, suggesting that stylistic them regardless of novelty—P8 described one response as “jargon variation is valued whether achieved through higher temperature or with too much complex words, very unpleasant for the reader,” while mild repulsion. Grouping by type, consensus-aware systems (SRTP14 dismissed another with “What??” Conversely, the desire for Mild + SRT-Strong) received 67% of all selections (14/21) vs 33% differentiation was strong: P3 stated “I want my responses to be diffor baselines (7/21). Standard baseline sampling was selected by ferent than those commonly generated,” and P12 asked “Why would only one participant (6%), suggesting that users value some form of I write a story that 500 other people already did!” This conditional divergence capability when given explicit choice. framing contextualizes the quantitative coherence finding (𝑝 = .006): Originality, Coherence, and Transparency. Originality ratings participants actively screened for readability, yet SRT-Strong scored favoured SRT-Strong (𝑀 = 3.40) over Baseline-Pure (𝑀 = 2.85, highest, suggesting its fluency controls maintained the baseline par+19%), though this difference approached but did not reach signifticipants demanded. Simultaneously, 8/16 expressed overcorrection icance, 𝜒 2 (3, 𝑁 = 16) = 7.34, 𝑝 = .062, 𝑊 = .153. This pattern— fears—P14 warned “If you max it out, the text would become trash significant practical utility (𝑝 = .019) and coherence (𝑝 = .006) to read. Not smooth at all,” and P15 worried the result would be alongside marginal perceived originality (𝑝 = .062)—suggests that “too different to my request.” These fears persisted even though the users reliably detect the usefulness of divergent outputs without requantitative data showed positive originality-coherence correlations liably detecting originality in blind evaluation. This is consistent (𝜌 = +.40 to +.67), suggesting participants hold intuitive mental with the broader framing of SRT as a consensus-avoidance tool: models where novelty costs readability—an expectation that did not awareness of consensus may matter more than perceptual sensitivity materialize within the tested 𝜆 range. to originality as such. Critically, originality and coherence ratings Theme 2: Strategic Rather Than Maximal Divergence (13/16). were positively correlated across all systems (𝜌 = +.40 to +.67), Participants overwhelmingly preferred context-dependent use of dicontradicting the expected readability-novelty trade-off and providvergence control. Eleven participants preferred context-dependent ing indirect evidence that divergence does not produce off-topic slider usage; only 3/16 indicated they would maximize divergence, wandering. Participants rated SRT’s transparency features highly with the remaining 2 preferring typical levels or finding the control (𝑀 = 3.62–4.38/5), with 68.8% indicating they would adopt such not yet useful. P8 stated: “I would like to have a control of the text features in future AI tools. divergence and use it according to project.” P5 described a multiConsensus Visualization. When asked to interpret the Yellow system workflow: “probably a mix of both—start with System B Zone before explanation, only 4/16 (25%) correctly identified it then D,” and P10 similarly wanted to “combine B and D.” Evaluation
Khan et al.
Figure 6: User study results demonstrating practical utility and adoption intent. (a) Perceived usefulness ratings across systems (Friedman 𝑝 = .019). (b) Willingness to use as a creative starting point, shown as percentage of participants expressing willingness per task, aggregated across all three tasks (Cochran’s Q 𝑝 = .075). (c) System preferences when given explicit choice; two participants selected “None.” Error bars show 95% confidence intervals. Systems: BP=Baseline-Pure, BH=Baseline-HighTemp, SM=SRT-Mild, SS=SRT-Strong (𝑁 = 16). criteria also shifted by task mode: in Technical writing, P12 noted “I do not consider the creativity matter in this topic,” while in Brainstorming, P8 valued ideas “usable in the real world scenario” and P15 praised responses that were “creative-practical”—suggesting users apply different evaluative standards that align with the system’s mode-specific configurations. Several participants conceptualized AI output as raw material for human editing rather than finished product: P16 explained wanting to “see the most frequent combinations first” and then “revise to make the writing more creative later,” while P8 stated a preference for editing “the prewritten text by myself.” Theme 3: Consensus Awareness (10/16). A majority reported that the study changed how they think about AI-generated content. P2 stated: “I never thought about consensus before, now I’ll look at responses from the lens of consensus.” P7 noted that “default settings clearly push people toward narrow conceptual and stylistic basins,” and P16 reported: “I will pay more attention to the homogenization content provided by AI.” Among the 6/16 who reported no shift, reasons included pre-existing awareness (P1, who already recognized homogenization as “regression to the mean,” saw no new shift) and general skepticism (P11). When asked directly whether AI homogenization is a real concern, 10/16 agreed, though depth varied: P12 framed it as “a natural outcome” rather than a design problem, while P4 argued effective prompting overcomes convergence. Theme 4: Divergence Skeptics (3/16). Not all participants embraced automated divergence. P1, a daily creative writer who rarely uses AI, rejected all systems (“All generated absolutely awful content”) and reported no change in thinking, representing a consistent stance that AI-generated text is unsuitable for creative work regardless of divergence settings. P16—the only participant to prefer System A (Standard AI)—wanted consensus visibility as information rather than as a trigger for automated divergence: “Choosing the creative AI systems can make my writing less human-like and I might put into more efforts.” P2 raised a saturation concern: “if I depend too much on AI for consensus avoidance. . . wouldn’t we reach saturation?” These disconfirming cases indicate that consensus-aware generation is not universally desired; some users prefer manual editing over automated divergence, and others question whether
systematic consensus-avoidance introduces its own limitations. We note that as system designers, our analytical lens may orient toward divergence-positive interpretations; we actively sought and report these disconfirming cases to counterbalance that tendency.
5
Discussion and Conclusion
SRT demonstrates that semantic divergence need not sacrifice coherence— a finding supported by both direct ratings and their relationship. SRTStrong achieved higher ratings for coherence (𝑀 = 3.92, 𝑝 = .006, 𝑊 = .260) and usefulness (𝑀 = 3.60, 𝑝 = .019, 𝑊 = .208), with the highest originality ratings (𝑀 = 3.40, 𝑝 = .062) among all systems. Originality and coherence ratings were positively correlated across all systems (𝜌 = +.40 to +.67), indicating that participants who perceived outputs as more original also tended to rate them as more coherent. This positive relationship also provides indirect evidence against mere semantic drift: if increased embedding distance reflected off-topic wandering rather than meaningful novelty, we would expect an inverse relationship with coherence ratings. Yet half of participants feared coherence trade-offs—an expectation-reality gap suggesting intuitive mental models where creativity costs readability, even when evidence shows otherwise within the tested 𝜆 range. The marginal significance of perceived originality (𝑝 = .062) alongside significant usefulness (𝑝 = .019) suggests users recognize practical value in divergent outputs without reliably detecting originality in blind evaluation. Awareness of consensus may matter more than perceptual sensitivity—pointing toward design opportunities in making the positive originality-coherence relationship explicit to users. Design Implications. Three qualitative findings carry design implications. First, only 25% correctly identified the Yellow Zone as consensus clustering, while 38% interpreted it as indicating optimal responses—revealing that users read visual density as quality rather than frequency. Future consensus visualizations should use explicit labeling (e.g., “common” vs. “good”) and progressive disclosure to prevent this conflation.
Seeing the Hivemind: A Consensus-Aware Interaction Technique for Mitigating AI Homogenization
Second, participants applied different evaluation criteria across task modes: accuracy in Technical writing, feasibility in Brainstorming, and vividness in Creative writing. This validates SRT’s modespecific configurations (different 𝜏, 𝜌, 𝑛 per mode) and suggests consensus-aware systems should surface mode-appropriate quality indicators alongside divergence metrics. Third, several participants wanted consensus visibility for manual editing rather than automated divergence—preferring to see frequent combinations first then revise themselves—suggesting SRT can also function as a diagnostic tool with users retaining editorial control. System preferences were elicited after the Yellow Zone explanation, which may have primed participants toward consensus-aware systems; these should be interpreted cautiously alongside blind Likert ratings. Limitations. Our sample (N=16) limits generalizability, and the visualization’s mixed interpretability (25% correct, 38% conflating density with quality) motivates progressive disclosure in future work. Ablation studies are needed to isolate component contributions, and longitudinal studies to assess whether consensus awareness durably changes creative practice.
References [1] Barrett R Anderson, Jash Hemant Shah, and Max Kreminski. 2024. Homogenization Effects of Large Language Models on Human Creative Ideation. In Proceedings of the 16th Conference on Creativity & Cognition (Chicago, IL, USA) (C&C ’24). Association for Computing Machinery, New York, NY, USA, 413–425. doi:10.1145/3635636.3656204 [2] Anthropic. 2024. Claude 3 Haiku: our fastest model yet. (2024). https://www. anthropic.com/news/claude-3-haiku Accessed: 2025-12-17. [3] Virginia Braun and Victoria Clarke. 2006. Using thematic analysis in psychology. Qualitative Research in Psychology 3, 2 (2006), 77–101. arXiv:https://doi.org/10.1191/1478088706qp063oa doi:10.1191/ 1478088706qp063oa [4] Ricardo Campos, Vítor Mangaravite, Arian Pasquali, Alípio Jorge, Célia Nunes, and Adam Jatowt. 2020. YAKE! Keyword extraction from single documents using multiple local features. Information Sciences 509 (2020), 257–289. [5] Liuqing Chen, Yaxuan Song, Chunyuan Zheng, Qianzhi Jing, Preben Hansen, and Lingyun Sun. 2025. Understanding Design Fixation in Generative AI. arXiv:2502.05870 [cs.HC] https://arxiv.org/abs/2502.05870 [6] Anil R. Doshi and Oliver P. Hauser. 2024. Generative AI enhances individual creativity but reduces the collective diversity of novel content. Science Advances 10, 28 (2024), eadn5290. arXiv:https://www.science.org/doi/pdf/10.1126/sciadv.adn5290 doi:10.1126/ sciadv.adn5290 [7] Katherine Fu, Joel Chan, Jonathan Cagan, Kenneth Kotovsky, Christian Schunn, and Kristin Wood. 2013. The meaning of “near” and “far”: the impact of structuring design databases and the effect of distance of analogy on design output. Journal of Mechanical Design 135, 2 (2013), 021007. [8] Xijin Ge. 2025. DataMap: A Portable Application for Visualizing HighDimensional Data. arXiv:2504.08875 [q-bio.QM] https://arxiv.org/abs/2504. 08875 [9] Kazjon Grace and Mary Lou Maher. 2019. Expectation-based models of novelty for evaluating computational creativity. In Computational creativity: The philosophy and engineering of autonomously creative systems. Springer, 195–209. [10] Nitin Gupta, Yoonhee Jang, Sara C Mednick, and David E Huber. 2012. The road not taken: Creative solutions require avoidance of high-frequency responses. Psychological science 23, 3 (2012), 288–294. [11] Ari Holtzman, Jan Buys, Li Du, Maxwell Forbes, and Yejin Choi. 2019. The curious case of neural text degeneration. arXiv preprint arXiv:1904.09751 (2019). [12] Liwei Jiang, Yuanjun Chai, Margaret Li, Mickel Liu, Raymond Fok, Nouha Dziri, Yulia Tsvetkov, Maarten Sap, Alon Albalak, and Yejin Choi. 2025. Artificial Hivemind: The Open-Ended Homogeneity of Language Models (and Beyond). arXiv:2510.22954 [cs.CL] https://arxiv.org/abs/2510.22954 [13] Jingoog Kim and Mary Lou Maher. 2023. The effect of AI-based inspiration on human design ideation. International Journal of Design Creativity and Innovation 11, 2 (2023), 81–98. [14] Taewook Kim, Matthew Kay, Yuqian Sun, Melissa Roemmele, Max Kreminski, and John Joon Young Chung. 2025. Scaffolding Recursive Divergence and Convergence in Story Ideation. arXiv:2507.03307 [cs.HC] https://arxiv.org/abs/
2507.03307 [15] Xiang Lisa Li, Ari Holtzman, Daniel Fried, Percy Liang, Jason Eisner, Tatsunori B Hashimoto, Luke Zettlemoyer, and Mike Lewis. 2023. Contrastive decoding: Open-ended text generation as optimization. In Proceedings of the 61st annual meeting of the association for computational linguistics (volume 1: Long papers). 12286–12312. [16] Alisa Liu, Maarten Sap, Ximing Lu, Swabha Swayamdipta, Chandra Bhagavatula, Noah A. Smith, and Yejin Choi. 2021. DExperts: Decoding-Time Controlled Text Generation with Experts and Anti-Experts. arXiv:2105.03023 [cs.CL] https: //arxiv.org/abs/2105.03023 [17] Shusen Liu, Dan Maljovec, Bei Wang, Peer-Timo Bremer, and Valerio Pascucci. 2016. Visualizing high-dimensional data: Advances in the past decade. IEEE transactions on visualization and computer graphics 23, 3 (2016), 1249–1268. [18] Leland McInnes, John Healy, and James Melville. 2018. Umap: Uniform manifold approximation and projection for dimension reduction. arXiv preprint arXiv:1802.03426 (2018). [19] Leland McInnes, John Healy, Nathaniel Saul, and Lukas Großberger. 2018. UMAP: Uniform Manifold Approximation and Projection. Journal of Open Source Software 3, 29 (2018), 861. doi:10.21105/joss.00861 [20] Meta. 2024. Llama 3. https://arxiv.org/abs/2407.21783 Accessed: 2025-12-17. [21] OpenAI. 2023. GPT-4 Technical Report. arXiv preprint arXiv:2303.08774 (2023). [22] Janet Rafner, Blanka Zana, Ida Bang Hansen, Simon Ceh, Jacob Sherson, Mathias Benedek, and Izabela Lebuda. 2025. Agency in Human-AI Collaboration for Image Generation and Creative Writing: Preliminary Insights from Think-Aloud Protocols. Creativity Research Journal 0, 0 (2025), 1–24. arXiv:https://doi.org/10.1080/10400419.2025.2587803 doi:10.1080/10400419. 2025.2587803 [23] Ben Shneiderman. 2007. Creativity support tools: accelerating discovery and innovation. Commun. ACM 50, 12 (Dec. 2007), 20–32. doi:10.1145/1323688. 1323689 [24] Taylor Sorensen, Jared Moore, Jillian Fisher, Mitchell Gordon, Niloofar Mireshghallah, Christopher Michael Rytting, Andre Ye, Liwei Jiang, Ximing Lu, Nouha Dziri, Tim Althoff, and Yejin Choi. 2024. A Roadmap to Pluralistic Alignment. arXiv:2402.05070 [cs.AI] https://arxiv.org/abs/2402.05070 [25] Sangho Suh, Meng Chen, Bryan Min, Toby Jia-Jun Li, and Haijun Xia. 2024. Luminate: Structured Generation and Exploration of Design Space with Large Language Models for Human-AI Co-Creation. In Proceedings of the 2024 CHI Conference on Human Factors in Computing Systems (Honolulu, HI, USA) (CHI ’24). Association for Computing Machinery, New York, NY, USA, Article 644, 26 pages. doi:10.1145/3613904.3642400 [26] Sean Welleck, Ilia Kulikov, Stephen Roller, Emily Dinan, Kyunghyun Cho, and Jason Weston. 2019. Neural Text Generation with Unlikelihood Training. arXiv:1908.04319 [cs.LG] https://arxiv.org/abs/1908.04319 [27] An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, Huan Lin, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jingren Zhou, Junyang Lin, Kai Dang, Keming Lu, Keqin Bao, Kexin Yang, Le Yu, Mei Li, Mingfeng Xue, Pei Zhang, Qin Zhu, Rui Men, Runji Lin, Tianhao Li, Tianyi Tang, Tingyu Xia, Xingzhang Ren, Xuancheng Ren, Yang Fan, Yang Su, Yichang Zhang, Yu Wan, Yuqiong Liu, Zeyu Cui, Zhenru Zhang, and Zihan Qiu. 2025. Qwen2.5 Technical Report. arXiv:2412.15115 [cs.CL] https://arxiv.org/abs/2412.15115 [28] Jiayi Zhang, Simon Yu, Derek Chong, Anthony Sicilia, Michael R. Tomz, Christopher D. Manning, and Weiyan Shi. 2025. Verbalized Sampling: How to Mitigate Mode Collapse and Unlock LLM Diversity. arXiv:2510.01171 [cs.CL] https://arxiv.org/abs/2510.01171
A
Experiment 2: Full Ablation Results
This appendix reports the complete 8-condition ablation results across all three modes, supplementing the main-effect delta summary in Table 2. The 23 factorial design toggles Contrastive Decoding (CD), Phrase Penalties (PP), and Fluency Controls (FL) independently, yielding eight conditions per mode. Conditions are labelled C0 (no mechanisms) through C7 (all mechanisms enabled); the full condition legend is given in Table 8. All values are means over 𝑁 = 5 samples × 10 prompts = 50 observations per cell. Figure 7 visualises the full ablation across all conditions and modes for Originality, Relevance, and Cliché Frequency.
Khan et al.
Figure 7: Full ablation: 8 conditions × 3 modes for Originality, Relevance, and Cliché Frequency. Highlighted bar (C7) = Full SRT. C1 and C4 (CD active, FL inactive) achieve high originality but at the cost of catastrophic coherence (PPL > 88,000), confirming that Fluency Controls are a necessary stabilizer rather than an optional component.
B
Experiment 3: Full 𝜆-Sweep Results
This appendix reports the complete per-mode 𝜆-sweep results supplementing the overall summary in Table 3. Values are means ±
SD over 𝑁 = 5 samples × 10 prompts = 50 observations per cell. The Relevance and Coherence (−NLL) metrics characterize the
Seeing the Hivemind: A Consensus-Aware Interaction Technique for Mitigating AI Homogenization
Table 5: Full ablation results — Creative mode (mean, 𝑁 = 50). Shaded row = Full SRT (C7). Cond. CD PP FL Orig.
Rel.
Cliché
PPL
C0 C1 C2 C3 C4 C5 C6 C7
0.434 0.040 0.420 0.473 0.039 0.298 0.415 0.299
1.380 0.000 0.140 1.280 0.000 0.180 0.140 0.040
5.9 123,079 6.9 5.5 119,932 75.9 5.9 71.3
× ✓ × × ✓ ✓ × ✓
× × ✓ × ✓ × ✓ ✓
× × × ✓ × ✓ ✓ ✓
0.293 0.925 0.317 0.258 0.921 0.482 0.305 0.484
Table 6: Full ablation results — Technical mode (mean, 𝑁 = 50). Shaded row = Full SRT (C7). Cond. CD PP FL Orig.
Rel.
Cliché
PPL
C0 C1 C2 C3 C4 C5 C6 C7
0.722 0.171 0.713 0.716 0.170 0.662 0.715 0.663
0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000
2.0 88,739 2.0 2.1 90,267 9.2 2.1 8.9
× ✓ × × ✓ ✓ × ✓
× × ✓ × ✓ × ✓ ✓
× × × ✓ × ✓ ✓ ✓
0.054 0.812 0.060 0.055 0.829 0.149 0.056 0.151
Table 7: Full ablation results — Brainstorming mode (mean, 𝑁 = 50). Shaded row = Full SRT (C7). Cond. CD PP FL Orig.
Rel.
Cliché
PPL
C0 C1 C2 C3 C4 C5 C6 C7
0.553 0.074 0.536 0.554 0.086 0.452 0.550 0.454
2.320 0.000 1.160 2.260 0.000 0.200 1.380 0.220
5.4 112,280 5.3 5.4 112,834 41.9 5.6 40.0
× ✓ × × ✓ ✓ × ✓
× × ✓ × ✓ × ✓ ✓
× × × ✓ × ✓ ✓ ✓
0.158 0.914 0.183 0.150 0.903 0.326 0.158 0.355
D
C0 C1 C2 C3 C4 C5 C6 C7
Metric Definitions
Statistical Significance: 𝜆-Sweep
Table 11 reports Wilcoxon signed-rank test results comparing each 𝜆 > 0 to the 𝜆 = 0 baseline, with Bonferroni correction for four simultaneous comparisons (𝛼 corrected = 0.0125). The Friedman test across all five 𝜆 values additionally confirms overall differences for each metric. All effects are 𝑝 < 0.001, confirming that even the smallest tested step (𝜆 = 0.6) produces statistically reliable divergence and that the observed trends are not attributable to sampling variance.
Table 8: Ablation condition legend. Condition
C
For reproducibility, we define all six metrics used across Experiments 1–3. Originality. Cosine distance between output embedding e𝑔 and consensus centroid c: 𝐷 mode (𝑔) = 1 − e𝑔⊤ c. Higher = more divergent from consensus. Embeddings are 384dimensional L2-normalised vectors from all-MiniLM-L6-v2. Intra-Diversity. Mean pairwise cosine distance between all outputs from the same system for the same prompt. Higher = more internally varied. Cliché Frequency. Count of consensus-characterizing phrases (extracted by YAKE [4] for Creative, curated lists for Technical and Brainstorming) appearing in each output. Lower = better consensus suppression. Relevance. Cosine similarity between the prompt embedding and the output embedding: sim(eprompt, eoutput ). Higher = more on-topic. Coherence (−NLL). Negative mean token log-likelihood of Í the output under the strong model: − |𝑦1 | 𝑡 log 𝑝𝜃 (𝑦𝑡 | 𝑦 <𝑡 ). Higher (less negative) = more fluent. Provides an objective, scalable coherence proxy independent of human raters. Perplexity. exp(NLL), i.e., exp(−Coherence). Lower = more fluent. A perplexity of 5 is characteristic of normal fluent text; values above 1,000 indicate near-incoherent output. Note on embedding-based metrics. Originality and Relevance are both computed from the same all-MiniLM-L6-v2 embeddings. These metrics capture semantic distance at the sentence level but cannot distinguish diversity of expression (different words, same idea) from diversity of ideas (genuinely different concepts). Future work should complement these metrics with expert annotation or structured idea-level evaluation.
Contrastive Decoding
Phrase Penalties
Fluency Controls
× ✓ × × ✓ ✓ × ✓
× × ✓ × ✓ × ✓ ✓
× × × ✓ × ✓ ✓ ✓
divergence–relevance–coherence trade-off across the full operating range. Figures 8 and 9 show all five metrics per mode across the full 𝜆 range. Figure 10 shows cliché suppression separately by mode.
Khan et al.
Table 9: Per-mode 𝜆-sweep results: mean ± SD (𝑁 = 50 per cell). Orig. = Originality; Div. = Intra-Diversity; Cliché = Consensus Phrase Frequency; Rel. = Relevance; Coh. = Coherence (−NLL); PPL = Perplexity. Mode
𝜆
Orig.
Div.
Cliché
Rel.
Coh.
PPL
Technical
0.0 0.6 1.2 1.8 2.4
0.096 ± 0.084 0.111 ± 0.086 0.179 ± 0.099 0.212 ± 0.101 0.215 ± 0.081
0.103 ± 0.069 0.127 ± 0.066 0.195 ± 0.098 0.201 ± 0.095 0.215 ± 0.089
0.000 ± 0.000 0.000 ± 0.000 0.000 ± 0.000 0.000 ± 0.000 0.000 ± 0.000
0.710 ± 0.082 0.696 ± 0.086 0.661 ± 0.094 0.653 ± 0.095 0.653 ± 0.096
−0.71 ± 0.23 −1.13 ± 0.46 −1.95 ± 0.93 −2.35 ± 1.10 −2.71 ± 1.17
2.1 ± 0.5 3.4 ± 1.7 10.2 ± 8.6 17.9 ± 17.6 26.5 ± 26.4
Creative
0.0 0.6 1.2 1.8 2.4
0.293 ± 0.088 0.390 ± 0.083 0.466 ± 0.099 0.509 ± 0.107 0.500 ± 0.103
0.404 ± 0.101 0.489 ± 0.090 0.499 ± 0.104 0.465 ± 0.093 0.424 ± 0.093
0.200 ± 0.404 0.080 ± 0.274 0.060 ± 0.240 0.060 ± 0.240 0.020 ± 0.141
0.417 ± 0.130 0.332 ± 0.120 0.310 ± 0.123 0.274 ± 0.127 0.288 ± 0.117
−1.79 ± 0.17 −3.01 ± 0.22 −4.26 ± 0.26 −4.73 ± 0.29 −4.97 ± 0.32
6.1 ± 1.0 20.9 ± 5.0 73.2 ± 19.3 118.2 ± 33.4 150.9 ± 46.3
0.0 0.6 Brainstorming 1.2 1.8 2.4
0.162 ± 0.069 0.257 ± 0.083 0.329 ± 0.092 0.374 ± 0.078 0.393 ± 0.093
0.242 ± 0.074 0.358 ± 0.051 0.399 ± 0.051 0.373 ± 0.087 0.415 ± 0.076
1.340 ± 1.272 0.260 ± 0.487 0.160 ± 0.370 0.080 ± 0.274 0.040 ± 0.198
0.551 ± 0.097 0.502 ± 0.103 0.472 ± 0.091 0.456 ± 0.082 0.434 ± 0.117
−1.60 ± 0.33 −2.52 ± 0.36 −3.63 ± 0.55 −4.31 ± 0.42 −4.49 ± 0.46
5.3 ± 2.0 13.2 ± 4.9 44.9 ± 35.1 80.5 ± 32.3 99.3 ± 48.0
Table 10: Marginal originality and relevance gains per 𝜆 step (averaged across modes, 𝑁 = 150). The near-zero gain at 1.8 → 2.4 confirms the practical operating ceiling. 𝜆 step
ΔOrig.
ΔRel.
ΔPPL
0.0 → 0.6 0.6 → 1.2 1.2 → 1.8 1.8 → 2.4
+0.069 +0.072 +0.040 +0.005
−0.049 −0.029 −0.020 −0.003
+8.1 +30.2 +29.5 +20.0
Table 11: Wilcoxon signed-rank tests: each 𝜆 vs. 𝜆 = 0 baseline (𝑁 = 150 per 𝜆, Bonferroni corrected). All comparisons: *** 𝑝 < 0.001. 𝜆 vs 0
Orig.
Rel.
Cliché
Coh.
0.6 vs 0.0 1.2 vs 0.0 1.8 vs 0.0 2.4 vs 0.0
*** *** *** ***
*** *** *** ***
*** *** *** ***
*** *** *** ***
Friedman 𝜒 2 p
338.67 < .001
161.54 < .001
114.69 < .001
525.60 < .001
Seeing the Hivemind: A Consensus-Aware Interaction Technique for Mitigating AI Homogenization
Figure 8: Per-mode 𝜆-sweep (Part 1 of 2): Originality (green, ↑ divergence), Intra-Diversity (blue, ↑ varied), and Cliché Frequency (orange, ↓ better suppression). Rows correspond to modes; columns to metrics. 𝑁 = 50 per mode per 𝜆; error bars = SD.
Khan et al.
Figure 9: Per-mode 𝜆-sweep (Part 2 of 2): Relevance (blue, ↑ on-topic) and Coherence −NLL (red, ↑ fluent). Technical mode (top row) retains relevance of 0.653 even at 𝜆 = 2.4, an 8.1% drop from baseline, validating its lower default configuration. Creative mode (middle row) shows the steepest relevance decline (30.9%), with the originality–relevance crossover at 𝜆 ≈ 0.8.
Seeing the Hivemind: A Consensus-Aware Interaction Technique for Mitigating AI Homogenization
Figure 10: Cliché suppression by mode. Brainstorming begins with the highest baseline cliché rate (1.34 at 𝜆 = 0) due to marketingframe prevalence in LLM brainstorming outputs, and reaches near-zero by 𝜆 = 1.2. Technical mode shows zero cliché frequency throughout because the curated boilerplate phrases were not present in these particular prompts’ outputs—a finding that motivates extending the Technical mode phrase list in future work.
Figure 11: Statistical significance heatmap. Color intensity encodes − log10 (𝑝); all cells show *** (𝑝 < 0.001), confirming that every 𝜆 step produces reliable, statistically significant effects across all four metrics.
Khan et al.
Algorithm 1 Semantic Repulsion Technique (SRT) 1.5em 1.1em Require: Prompt 𝑃 , task mode 𝑀 , repulsion strength 𝜆 Require: Strong model 𝑆 , weak model 𝑊 , embedding model 𝐸 Ensure: Repulsed response 𝑌
1: X ← ∅
31: end if // Prepare decoding
32: 𝑉 ← A LIGN VOCABULARIES (𝑆,𝑊 ) 33: 𝑌 ← TOKENIZE (𝑃 ) // Generate using semantic repulsion
// Generate default responses to estimate consensus
2: for 𝑖 = 1 to 12 do 3: 𝑥𝑖 ← G ENERATE (𝑆, 𝑃, temperature = 1.0, top-𝑝 = 0.9) 4: X ← X ∪ {𝑥𝑖 } 5: end for 6: X ← R EMOVE D UPLICATES ( X) // Compute semantic consensus centroid
7: for all 𝑥𝑖 ∈ X do 8: 𝑒𝑖 ← N ORMALIZE (𝐸 (𝑥𝑖 ) ) 9: end for Í 10: 𝑐 ← N ORMALIZE ( 𝑖 𝑒𝑖 ) // Extract consensus phrases to avoid
11: if 𝑀 = Creative then 12: N ← E XTRACT R EPEATED K EYPHRASES ( X, 𝑃 ) 13: else if 𝑀 = Technical then 14: N ← F IND B OILERPLATE P HRASES ( X) 15: else if 𝑀 = Brainstorm then 16: N ← F IND M ARKETING C LICHES ( X) 17: if | N | < 12 then 18: N ← N ∪ E XTRACT R EPEATED K EYPHRASES ( X, 𝑃 ) 19: end if 20: end if // Convert negative phrases into token penalties
21: Tneg ← ∅ 22: for all 𝑛 ∈ N do 23: Tneg ← Tneg ∪ T OKENIZE (𝑛) 24: end for // Set mode-specific generation parameters 25: if 𝑀 = Creative then 26: 𝜏 ← 0.003, 𝛽 ← 4.0, 𝜌 ← 1.06, 𝑔 ← 4, 𝐿 ← 320 27: else if 𝑀 = Technical then 28: 𝜏 ← 0.010, 𝛽 ← 4.0, 𝜌 ← 1.12, 𝑔 ← 6, 𝐿 ← 220 29: else if 𝑀 = Brainstorm then 30: 𝜏 ← 0.002, 𝛽 ← 4.0, 𝜌 ← 1.06, 𝑔 ← 4, 𝐿 ← 260
34: for 𝑡 = 1 to 𝐿 do 35: ℓ𝑆 ← 𝑆. N EXT T OKEN L OGITS (𝑌 ) 36: 𝑌𝑊 ← T OKENIZE𝑊 ( D ETOKENIZE (𝑌 ) ) raw ← 𝑊 . N EXT T OKEN L OGITS (𝑌 ) 37: ℓ𝑊 𝑊 raw , 𝑉 ) 38: ℓ𝑊 ← A LIGN L OGITS (ℓ𝑊 // Core semantic repulsion step
39: 40: 41: 42:
ℓ ← ℓ𝑆 − 𝜆ℓ𝑊 // Penalize consensus tokens for all 𝑣 ∈ Tneg do
ℓ [𝑣 ] ← ℓ [𝑣 ] − 𝛽
end for // Preserve fluency 43: 𝑝𝑆 ← S OFTMAX (ℓ𝑆 ) 44: for all token 𝑣 do 45: if 𝑝𝑆 [𝑣 ] < 𝜏 then
46: 47: 48:
ℓ [𝑣 ] ← −∞
end if end for // Avoid repetition 49: ℓ ← A PPLY R EPETITION P ENALTY (ℓ, 𝑌 , 𝜌 ) 50: ℓ ← B LOCK R EPEATED N GRAMS (ℓ, 𝑌 , 𝑔) // Sample next token 51: ℓ ← T OP PF ILTER (ℓ, 0.9) 52: 𝑦𝑡 ← S AMPLE ( S OFTMAX (ℓ ) )
53: 𝑌 ← 𝑌 ∪ {𝑦𝑡 } 54: if 𝑦𝑡 = E ND O F T EXT then 55: break 56: end if 57: end for 58: 𝑌 ← D ETOKENIZE (𝑌 ) 59: 𝑌 ← R EMOVE P ROMPT P REFIX (𝑌 , 𝑃 ) 60: return 𝑌