PaperlayerView exact revision on arXiv

Source-linked AI summary

Artificial Hivemind: The Open-Ended Homogeneity of Language Models (and Beyond)

Liwei Jiang, Yuanjun Chai, Margaret Li, Mickel Liu, Raymond Fok, Nouha Dziri, Yulia Tsvetkov, Maarten Sap, Alon Albalak, Yejin Choi

arXiv:2510.22954v1cs.CL

TL;DR

Real-world open-ended LM diversity lacks scalable evaluation, limiting understanding of potential homogenization in model outputs. The paper introduces INFINITY-CHAT to study this gap and finds strong intra-model repetition and inter-model homogeneity, alongside miscalibration to divergent human preferences.

  • Problem

    Scalable evaluations of LM diversity in real-world, open-ended interactions remain limited beyond narrow, stylized tasks, despite concerns about homogenized outputs.

    Source: §1 · pp. 1–2

  • Method

    The paper constructs INFINITY-CHAT from 26,070 real-world open-ended queries, develops a 6-category, 17-subcategory taxonomy, and collects dense human ratings.

    Source: §1 · p. 2, §2 · p. 3, §4.1 · p. 7

  • Results

    Across 70+ LMs, the study finds pronounced intra-model repetition and inter-model homogeneity, while models and judges are miscalibrated to divergent human preferences.

    Source: §1 · p. 2

  • Takeaways & Limitations

    INFINITY-CHAT provides a foundation for diagnosing, benchmarking, and mitigating mode collapse and for studying diversity in model outputs.

    Source: §6 · p. 10

    Decoding-based approaches are less effective for close-ended tasks and place strategy-selection burdens on users, motivating more generalizable training-level solutions.

    Source: §3 · p. 5

Abstract

from arXiv · show

Language models (LMs) often struggle to generate diverse, human-like creative content, raising concerns about the long-term homogenization of human thought through repeated exposure to similar outputs. Yet scalable methods for evaluating LM output diversity remain limited, especially beyond narrow tasks such as random number or name generation, or beyond repeated sampling from a single model. We introduce Infinity-Chat, a large-scale dataset of 26K diverse, real-world, open-ended user queries that admit a wide range of plausible answers with no single ground truth. We introduce the first comprehensive taxonomy for characterizing the full spectrum of open-ended prompts posed to LMs, comprising 6 top-level categories (e.g., brainstorm & ideation) that further breaks down to 17 subcategories. Using Infinity-Chat, we present a large-scale study of mode collapse in LMs, revealing a pronounced Artificial Hivemind effect in open-ended generation of LMs, characterized by (1) intra-model repetition, where a single model consistently generates similar responses, and more so (2) inter-model homogeneity, where different models produce strikingly similar outputs. Infinity-Chat also includes 31,250 human annotations, across absolute ratings and pairwise preferences, with 25 independent human annotations per example. This enables studying collective and individual-specific human preferences in response to open-ended queries. Our findings show that LMs, reward models, and LM judges are less well calibrated to human ratings on model generations that elicit differing idiosyncratic annotator preferences, despite maintaining comparable overall quality. Overall, INFINITY-CHAT presents the first large-scale resource for systematically studying real-world open-ended queries to LMs, revealing critical insights to guide future research for mitigating long-term AI safety risks posed by the Artificial Hivemind.

1 Introduction

INFINITY-CHAT addresses the lack of scalable, realistic evaluations of open-ended LM diversity by combining 26K real-world queries, a 6-category/17-subcategory taxonomy, and dense human annotations. The study reveals Artificial Hivemind behavior within and across models and miscalibration toward divergent human preferences.

Motivation

  • Existing diversity benchmarks remain narrow, targeting stylized tasks such as persona generation, keyword-driven storytelling, random numbers, poetry, or figurative language rather than pluralistic real-world interactions.Although ensemble methods and model swarms have been proposed, scalable evaluations in realistic settings remain lacking.Source: §1 · pp. 1–2

Dataset and taxonomy

  • INFINITY-CHAT contains 26K real-world open-ended queries with no single correct answer and introduces a taxonomy of 6 top-level categories and 17 subcategories.The queries were mined from WildChat and include types such as Brainstorm & Ideation, Speculative & Hypothetical Scenarios, and Skill Development.Source: §1 · p. 2

Artificial Hivemind

  • Across 70+ open- and closed-source LMs, the study finds intra-model repetition and more critically inter-model homogeneity, as models converge on similar ideas with minor phrasing changes.This Artificial Hivemind effect suggests ensembles may lack true diversity when models share alignment and training priors.Source: §1 · p. 2

Human evaluation

  • 31,250 annotations, including absolute ratings and pairwise preferences from 25 independent annotators per query–response pair, support analysis of pluralistic human judgments.The annotations cover distinct model responses in INFINITY-CHAT.Source: §1 · p. 2
  • LMs, reward models, and LM judges are often miscalibrated on responses that elicit divergent annotator preferences despite comparable overall quality.This reflects modeling pipelines that assume a single consensus notion of quality and overlook diverse preferences.Source: §1 · p. 2

Overall contribution

  • The framework integrates realistic queries, query-type taxonomy, and dense human annotations to diagnose Artificial Hivemind behavior and guide safer, more expressive, and resourceful LMs.Its broader aim is to support models that better empower human creativity.Source: §1 · p. 2

2 INFINITY-CHAT: Real-World Open-Ended Queries with Diverse Responses

INFINITY-CHAT is a large-scale dataset of real-world open-ended user queries designed to study diverse language-model responses. It provides a comprehensive taxonomy of query types and exposes both the breadth of users’ open-ended requests and substantial within-model response repetition.

Dataset construction

  • INFINITY-CHAT contains 26,070 open-ended and 8,817 closed-ended queries mined from 37,426 filtered, high-quality GPT-4 user queries.Queries were English, non-toxic, single-turn, 15–200 characters, and refined for clarity when ambiguous.Source: §2 · p. 3

Query taxonomy

  • The taxonomy organizes real-world open-ended queries into 6 high-level categories and 17 fine-grained subcategories through iterative manual refinement and GPT-4o annotation.The resulting taxonomy is designed to capture the diverse landscape of user queries that invite multiple valid responses.Source: §2 · p. 3, Figure 2 · p. 3

Query distribution

  • Creative Content Generation dominates at 58.0%, while Alternative Writing Genres, Concept Explanation, Skill Development, Analytical & Interpretive Questions, and Hypothetical Scenarios each account for 22.2%–38.5% of queries.Brainstorming & Ideation comprises 15.2% of queries, highlighting users’ reliance on LMs for ideas and inspirations.Source: §2 · p. 3

Diversity resource

  • For 50 responses per query across 100 open-ended queries, average pairwise sentence-embedding similarity typically exceeds 0.8, whereas globally random pairs all fall in the 0.1–0.2 range.This resource supports studying LM output variation and advancing pluralistic alignment.Source: Figure 4 · p. 4, §2 · p. 4

Novel categories

  • The mining process identifies 314 novel open-ended categories, including prominent dimensions such as Cultural, Analysis, Ethical, Historical, Media, and Humor.These categories extend beyond the predefined taxonomy and reveal underexplored aspects of real-world open-ended use.Source: Figure 3 · p. 3, §2 · p. 4

3 Artificial Hivemind: Intra- and Inter-Model Homogeneity in LMs

Across 70+ language models and real-world open-ended queries, the study finds severe intra-model repetition and inter-model homogeneity. Similarity persists under aggressive and diversity-oriented decoding, appearing as semantic convergence, verbatim overlap, and cross-model clusters.

Intra-model repetition

  • 79% of query-level response pools had average pairwise similarity above 0.8, even with top-p = 0.9 and temperature = 1.0 sampling.The result uses 50 responses per query from INFINITY-CHAT100 and measures average pairwise embedding similarity within each model’s response pool.Source: §3 · p. 4, Figure 5 · p. 4
  • 81% of min-p response pairs exceeded 0.7 similarity and 61.2% exceeded 0.8, showing mode collapse despite diversity-oriented decoding.Min-p reduced the most extreme repetition, but substantial similarity remained under top-p = 1.0, min-p = 0.1, and temperature = 2.0.Source: Figure 5 · p. 4, §3 · p. 5

Inter-model homogeneity

  • Cross-model response similarity ranged from 71% to 82%, including 0.82 for DeepSeek-V3 versus qwen-max-2025-01-25 and 0.81 versus gpt-4o-2024-11-20.GPT and Qwen API models also tended to show higher similarity with models outside their own families, although the causes remain unclear.Source: §3 · p. 5, Figure 6 · p. 5

Instance-level convergence

  • Models produced verbatim phrase overlaps and occasionally identical answers across families, demonstrating surface-level repetition even for fully open-ended queries.Examples include shared marketing phrases between DeepSeek-V3 and gpt-4o-2024-11-20 and an identical motto from qwen-max-2025-01-25 and qwen-plus-2025-01-25.Source: Figure 6 · p. 5, §3 · p. 6

Semantic convergence

  • For “Write a metaphor about time,” 50 responses from each of 25 models formed just two semantic clusters centered on “time is a river” and “time is a weaver.”This shows convergence at the level of core concepts, not only wording.Source: §3 · p. 6

Inter-model homogeneity

  • With N = 50, the most similar response clusters averaged ∼8 unique source models, and some queries exceeded 10, indicating frequent cross-model convergence.Distinct models sometimes generated responses more similar to one another than responses from the same model.Source: §3 · p. 6, Figure 8 · p. 6

4 How Do LMs, Reward Models, and LM Judges Handle Alternative Responses to Open-Ended Queries?

Dense annotations reveal substantial human disagreement over alternative responses to open-ended queries, while LMs, reward models, and LM judges align less reliably with human ratings when responses are similarly good or preferences diverge. These findings motivate more nuanced modeling of fine-grained quality distinctions and idiosyncratic human preferences.

4.1 Gathering Distributional Annotations Across Many Humans

  • The annotation study collected 18,750 absolute-rating labels and 12,500 pairwise-preference labels, with 25 annotations per example, forming a densely annotated dataset of alternative responses.Entropy distributions show frequent annotator disagreement, including response pairs with near-uniform support across preference options.Source: §4.1 · p. 7

4.2 Gathering LMs, Reward Models, and LM Judges Ratings

  • The study compares response perplexity, standardized reward-model outputs, and LM-judge ratings using overall-quality and HHH rubrics against human annotations.The evaluation covers 56 state-of-the-art LMs, 6 top-ranked reward models, and 4 LM judges.Source: §4.2 · p. 8

4.3 Comparing Model Ratings to Human Scores for Responses to Open-Ended Queries

  • Average human ratings are used as shared-judgment estimates because different annotators may prefer different responses that receive comparable overall approval.The comparison tests whether model scores correlate less reliably with responses humans broadly consider comparably good.Source: §4 · p. 7, §4.3 · p. 8
  • Correlations with human ratings drop substantially on similar-quality responses across absolute and pairwise setups, consistently across alternative subset-selection methods.The result highlights the need to model fine-grained distinctions among responses that humans consider equally high-quality.Source: Figure 10 · p. 8, §4.3 · p. 9, Figure 11 · p. 9
  • Correlations with human ratings also drop substantially on examples with high annotator disagreement across both absolute and pairwise setups, consistently across alternative grouping methods.This indicates that model ratings do not adequately capture idiosyncratic disagreement over the broad range of acceptable open-ended responses.Source: Figure 10 · p. 8, Figure 11 · p. 9, §4.3 · p. 10

5 Related Work

Prior work frames LM diversity collapse as a threat to creativity, pluralistic alignment, and trustworthiness, while developing specialized creativity evaluations and pluralistic-alignment methods. However, existing approaches often rely on predefined diversity dimensions, motivating broader evaluation and individual-level alignment.

The diversity collapse problem of LMs

  • Diversity collapse limits LM output variety and can reduce creativity, minority representation, bias resistance, utility, and trustworthiness [93] [75] [23] [52] [5] [38].Proposed mitigations include diversifying training corpora [84] [65], modifying training algorithms [92, 53], and using alternative decoding [82] [60] or prompting strategies.Source: §5 · p. 10

Measuring the creativity and divergent thinking of language models

  • Creativity research adapts psychometric tests, human evaluation, LLM judges, specialized benchmarks, and new metrics, yet LM creative outputs can remain homogeneous despite high individual creativity scores [16] [13] [45] [69] [54].The paper addresses this gap through a large-scale study of real-world open-ended queries, a comprehensive taxonomy, a query dataset, and dense human annotations.Source: §5 · p. 10

Disagreement and pluralistic alignment of language models

  • Pluralistic alignment seeks to serve varied populations through new methods, benchmarks, data strategies, multi-LM interaction, and cultural-diversity research, but many approaches still use predefined dimensions such as demographics, personality, or culture [78, 2, 19, 60, 48, 14, 26, 80, 12, 42, 73, 81, 17, 57, 68, 74, 18, 58, 51, 47, 55, 50, 35, 95, 63, 89, 22, 4].Individual-level alignment is proposed to support personal variation without relying on stereotypes.Source: §5 · p. 10

6 Conclusion

INFINITY-CHAT is introduced as a large-scale resource for evaluating LM diversity in naturally occurring, open-ended settings, revealing an Artificial Hivemind effect involving intra-model repetition and inter-model homogeneity. Its diverse prompt taxonomy and dense human preference annotations support diagnosing, benchmarking, and mitigating mode collapse and output homogenization.

6 Conclusion

  • INFINITY-CHAT evaluates LM diversity in naturally occurring, open-ended settings and reveals an Artificial Hivemind effect comprising intra-model repetition and inter-model homogeneity.The resource is designed to study diversity beyond narrow or artificial generation settings.Source: §6 · p. 10
  • Its diverse prompt taxonomy and dense human preference annotations provide a foundation for diagnosing, benchmarking, and mitigating mode collapse in generative AI.The authors position INFINITY-CHAT as infrastructure for future research on genuine output diversity and protecting human expression from homogenization.Source: §6 · p. 10

Appendices … A.3 Broader Implications

The appendices frame Artificial Hivemind as a diagnostic and mitigation foundation while highlighting limitations in coverage, evaluation, and causal explanation. They also warn that cross-model convergence may amplify homogenization in synthetic data and human expression, with broader ethical and societal consequences.

A.1 Limitations

  • The benchmark covers 26K queries but remains an English-language snapshot that may underrepresent linguistic, cultural, regional, and contextual creative diversity.Its WildChat-derived scope limits generalizability to non-English contexts, while the taxonomy is intended to support future multilingual adaptation.Source: §A.1 · p. 19
  • GPT-4o taxonomy labels achieved 74.7% human-annotator agreement, while semantic embedding similarity may not fully express multidimensional creative variation.The authors consider the labeling accuracy sufficient for representing category distributions but acknowledge room for improvement.Source: §A.1 · p. 19
  • The study emphasizes diversity under fixed decoding without studying quality, finds evaluators struggle to distinguish equal-quality responses, and does not establish the causes of homogenization.Future work should develop reliable automatic evaluation and test whether shared data, memorization, generalization, or other factors drive convergence.Source: §A.1 · p. 19

A.2 Future Directions

  • Future foundation-model studies will disentangle pretraining from post-training and quantify how supervised fine-tuning, RLHF/RLAIF, and constitutional training contribute to homogenized responses.The planned testbed extends beyond instruction-following models to isolate these training effects.Source: §A.2 · p. 20
  • The dataset and Artificial Hivemind metric can support diversity-aware training, decoding, evaluation, red-teaming, reinforcement learning, and curricula targeting prompts susceptible to mode collapse.Proposed strategies include rewarding multiple valid modes, benchmarking diverse beam search and nucleus-sampling variants, and gradually increasing prompt openness.Source: §A.2 · p. 20

A.3 Broader Implications

  • Model convergence may propagate into human writing, creative ideation, and divergent thinking, amplifying uniformity in linguistic and cognitive patterns at scale [46] [6] [7].The work provides quantitative tools for assessing long-term cultural and epistemic consequences of LLM-mediated communication.Source: §A.3 · p. 20
  • Single-model synthetic-data distillation can intensify mode collapse, while model swarms and multi-agent frameworks seek diversity by aggregating outputs from multiple models [3].The findings show that multiple models do not guarantee diversity in open-ended tasks.Source: §A.3 · p. 20
  • Distinct state-of-the-art models often produce highly similar open-ended answers, undermining assumed multi-model distillation diversity without Artificial Hivemind diagnostics.This risks perpetuating homogenization across generations of language models and affects synthetic-data curation, ensemble alignment, and constitutional AI.Source: §A.3 · p. 20

B INFINITY-CHAT: A Dataset of In-The-Wild Open-Ended User Queries · B.1 Details of the Mining of In-The-Wild Open-Ended User Queries · B.2 Details of Defining the Open-Ended Query Taxonomy

INFINITY-CHAT is constructed by mining real-world user queries and filtering them into open-ended and closed-ended subsets. Its open-ended queries are further organized through an automatically mined taxonomy using GPT-4o-2024-11-20.

B INFINITY-CHAT: A Dataset of In-The-Wild Open-Ended User Queries

B.1 Details of the Mining of In-The-Wild Open-Ended User Queries

  • The queries are curated from WILDCHAT, which contains diverse real-world conversations from in-the-wild users.Source: §B.1 · p. 21
  • Filtering for English, non-toxic and non-harmful, GPT-4-directed, and 15–200-character queries yields 37,426 candidates.Source: §B.1 · p. 21
  • GPT-4o-2024-11-20 classifies candidates by meaningful information, greeting or model inquiry, and whether responses should be single or multiple.Queries are revised during classification when unclear or ambiguous.Source: §B.1 · p. 21
  • Automatic filtering produces 26,070 open-ended and 8,817 closed-ended queries, defining INFINITY-CHAT as a dataset designed to elicit diverse, high-quality responses.Examples appear in Tables 3 and 4, while classification prompts appear in Figures 12 and 13.Source: §B.1 · p. 21

B.2 Details of Defining the Open-Ended Query Taxonomy

  • The open-endedness taxonomy is illustrated with additional examples, and GPT-4o-2024-11-20 automatically mines and annotates its categories.Table 2 provides examples, while Figure 14 shows the annotation prompt for new and existing categories.Source: §B.2 · p. 21

B.3 Human Validation of the Open-endedness of Queries in INFINITY-CHAT · C Artificial Hiveminds: Examining the Intra- and Inter- Model Homogeneity

Human validation confirms that INFINITY-CHAT queries are broadly open-ended, while analysis across 42 models finds substantial response homogeneity even for highly open-ended prompts. The paper further examines intra- and inter-model homogeneity across open- and closed-source language models.

B.3 Human Validation of the Open-endedness of Queries in INFINITY-CHAT

  • The validation asked whether each query supported multiple similarly valid answers and, when applicable, how many reasonable alternatives could apply.Answer-count options ranged from fewer than 3 to more than 20 alternatives.Source: §B.3 · pp. 21–22
  • By majority vote, 89% of sampled queries were judged open-ended, and an inclusive at-least-one-annotator criterion classified 100% as open-ended.The study sampled 100 evaluation examples and assigned three participants to each query.Source: §B.3 · p. 21
  • 81.27% of queries allowed at least three reasonable alternatives, while 34.66% allowed more than 20, showing that queries were rarely narrow or deterministic.These ratings motivate diversity-aware evaluation beyond benchmarks with limited answer ranges.Source: §B.3 · p. 22, Table 1 · p. 22
  • Annotators’ estimates of open-endedness varied considerably across individuals, making definitive per-query ratings challenging.The study therefore used a coarse grouping: highly open-ended if any annotator selected “More than 20 responses,” and low otherwise.Source: §B.3 · p. 22

C Artificial Hiveminds: Examining the Intra- and Inter- Model Homogeneity

  • Average sentence similarity was 0.800 for highly open-ended prompts versus 0.837 for low-open-endedness prompts across 42 models, indicating limited diversity gains.Prompts were classified as highly open-ended when any annotator selected “More than 20 responses.”Source: §B.3 · p. 22
  • The study analyzed intra- and inter-model homogeneity across a broad range of open-source and closed-source language models.This establishes the scope of the Artificial Hiveminds analysis beyond a single model family.Source: C Artificial Hiveminds: Examining the… · p. 29

C.1 Evaluation Setups · C.2 Intra-Model Homogeneity · C.3 Inter-Model Homogeneity

The evaluation uses a unified generation protocol across intra- and inter-model analyses, based on 100 representative open-ended prompts and standardized decoding and similarity computation. Extended results are provided for intra-model repetition and inter-model homogeneity across the curated model set.

C.1 Evaluation Setups

  • The study reuses one unified generation protocol and identical model outputs for both intra-model and inter-model analyses.The seed set contains 100 representative prompts from INFINITY-CHAT100, while the analyzed model list and main-paper subset are documented in Table 5.Source: §C.1 · p. 29
  • Each model generates 50 responses per prompt under both top-p and minimum-p decoding configurations.Top-p uses p = 0.9 and temperature = 1.0; minimum-p uses p = 1.0, min-p = 0.1, and temperature = 2.0, with a 2048-token limit.Source: §C.1 · p. 29
  • The evaluation spans local, API-accessed, and TogetherAI-hosted models while keeping decoding configurations consistent across generation methods.Similarities are computed as cosine similarity between response sentence embeddings produced by text-embedding-3-small.Source: §C.1 · p. 29
  • Tables 12–16 contain concrete examples of response pairs alongside their corresponding similarity scores.These examples are referenced as supplementary material for interpreting the similarity-based analyses.Source: §C.1 · p. 29

C.2 Intra-Model Homogeneity

  • Table 6 provides extended intra-model repetition results for every model listed in Table 5, complementing the results summarized in Figure 6.The passage identifies Table 6 as the detailed counterpart to the Figure 6 summary but does not report individual result values.Source: §C.2 · p. 29

C.3 Inter-Model Homogeneity

  • Tables 7–11 provide extended inter-model homogeneity results for every model listed in Table 5, complementing the results summarized in Figure 6.The supplied passage identifies the tables’ scope but does not provide their individual comparisons or values.Source: §C.3 · p. 29

C.4 Examining How Paraphrased Queries Influence Response Homogeneity … D.2 Model Selection and Score Generation Details

Across 42 models, paraphrased prompts still produced highly similar responses, while the annotation resource captured broad human judgments for evaluating whether model-based scores recognize multiple valid answers. The evaluation compares language-model, reward-model, and judge scores with average human ratings using correlations tailored to open-ended quality.

C.4 Examining How Paraphrased Queries Influence Response Homogeneity

  • 0.781 cross-paraphrase similarity versus 0.821 within-prompt similarity shows that models remain highly consistent even when prompts are paraphrased, differing by only 0.04.The experiment used 30 prompts, four paraphrases per prompt, 20 responses, and 42 models with top-p=0.9 and temperature=1.Source: §C.4 · pp. 29–30

D.1 Human Annotation Details

  • Clustering model outputs into 15 groups per query was used to select diverse, representative responses for human annotation rather than sampling repetitive outputs directly.The procedure sampled 50 queries, collected 50 responses per model, and selected one response from each cluster.Source: §D.1 · p. 49
  • 31,250 human annotations were collected across absolute ratings and pairwise preferences, with 25 annotations per example, enabling analysis of distributional preferences when multiple responses are valid.The dataset includes 18,750 absolute-rating labels and 12,500 pairwise-comparison annotations from 50 prompts.Source: §D.1 · p. 49
  • Human rating distributions varied substantially across examples, as measured by Shannon entropy over 25 absolute ratings per query-response pair.The entropy analysis complements analogous pairwise-preference results in the main paper.Source: §D.1 · p. 49

D.2 Model Selection and Score Generation Details

  • Model-based evaluation covers perplexity-based language-model scores, scalar reward-model scores, and prompting-based LM judges, with Pearson correlations against average human ratings.Lower perplexity is treated as higher linguistic plausibility, while higher reward-model scores indicate better predicted quality.Source: §D.2 · pp. 49–50
  • LM-judge evaluation uses off-the-shelf GPT-4o and Prometheus [40] under Overall and Helpful, Honest, Harmless rubrics, producing scores on a 1-to-5 scale.The resulting raw scores are correlated with corresponding average human ratings.Source: §D.2 · p. 50
  • Average human ratings represent shared judgment across annotator subpopulations, motivating evaluation of whether scoring models treat multiple similarly good responses as equally valid.The design uses 25 ratings per example to capture individual preferences while retaining a collective-quality signal.Source: §D.2 · p. 50

D.3 Similar-Quality Responses · D.4 Disagreed Responses · E NeurIPS Paper Checklist

Appendix analyses test whether the paper’s calibration findings are robust to alternative ways of selecting similar-quality and highly disagreed responses. The NeurIPS checklist affirms that the abstract and introduction accurately reflect the paper’s contributions and scope.

D.3 Similar-Quality Responses

  • Alternative similar-quality subset selection methods yield findings consistent with the Tukey-based analysis, strengthening the robustness of the conclusions.Table 19 reports Spearman correlations across similarity-based subsets for human scores, LM perplexities, reward-model scores, and LM-judge scores.Source: §D.3 · pp. 50–51, Table 19 · p. 51
  • The similar-quality analysis compares optimized sliding, centroid-based, distance-based, Tukey, median-expansion, and gap-based subset-selection procedures.These methods respectively target compact ranges, low variance, low pairwise distance, outlier-robust clustering, median-centered selection, or small consecutive gaps.Source: §D.3 · p. 51

D.3 Similar-Quality Responses · D.4 Disagreed Responses

  • Model-level breakdowns extend the similar-quality and disagreement analyses across all models for both absolute-rating and pairwise-preference tasks.Tables 23 and 25 cover similar human scores, while Tables 24 and 26 cover high human disagreement in the two rating settings.Source: §D.3 · p. 51, §D.4 · p. 52, Table 23 · p. 60

D.4 Disagreed Responses

  • Alternative disagreed-response subset selection methods likewise produce findings consistent with the main analysis, further supporting robustness across disagreement definitions.Table 20 reports Spearman correlations between human scores and LM, reward-model, and LM-judge outputs across highly disagreed subsets.Source: §D.4 · p. 52, Table 20 · p. 52
  • The disagreement analysis operationalizes annotator disagreement using entropy, grouped entropy, Gini impurity, pairwise disagreement, majority-versus-minority deviation, and single-label Fleiss kappa.The measures capture either label uncertainty, disagreement frequency, deviation from consensus, or chance-corrected agreement.Source: §D.4 · p. 52

E NeurIPS Paper Checklist

  • The NeurIPS checklist answers “Yes” that the abstract and introduction accurately reflect the paper’s contributions and scope.Its guidance requires claims to match theoretical and experimental results and to reflect their expected generalizability, assumptions, and limitations.Source: E NeurIPS Paper Checklist · p. 64

2. Limitations

The paper discusses its limitations and broader impacts in Appendix A, while reporting no theoretical results. It also provides reproducibility resources, experimental details, statistical analyses, compute information, and responsible-release documentation.

Limitations and broader impact

  • The authors provide a thorough limitation discussion and address broader societal impacts and responsible-release safeguards in Appendix A.They also confirm conformity with the NeurIPS Code of Ethics.Source: §2 · pp. 64, 67, 68

Theory assumptions and proofs

  • The paper has no theoretical results, so assumptions and complete proofs are not applicable.Source: §2 · pp. 64–65

Experimental reproducibility and open access

  • The authors state that Appendix B, C, and D contain the information needed to reproduce the main experiments and that all code and data are openly accessible.The paper also includes instructions for running the code and links to the dataset and code in the abstract.Source: §2 · pp. 65–66

Experimental setting and significance

  • Appendices B, C, and D provide all experimental details, while Appendix D reports statistical-significance analyses of Pearson correlation differences.Source: §2 · p. 66

Experiment compute resources

  • The appendix documents compute and human-annotation resources used for query mining, model-response generation, and human-label collection.Source: §2 · p. 67

Assets, human subjects, and LLM usage

  • The paper documents assets, human-annotation procedures, the absence of IRB requirements for innocuous annotations, and its use of LLMs for mining open-ended data.Source: §2 · pp. 69–70

Related papers

Loading 2510.22954v1…