PaperlayerView exact revision on arXiv

Source-linked AI summary

Big-Big-SycoBench: A Benchmark to Detect Sycophancy Through Model Responses to Valid vs. Invalid Arguments

Marthe Ballon, Arne Vanhoyweghen, Seorin Kim, Francesca Carlon, Lynn Delcon, Melika Mobini, Vincent Holst, Lukas Van de Voorde, Tibo Vanleke, Mathieu Louis, Bert Verbruggen, Brecht F. Verbeken, Andres Algaba, Vincent Ginis

arXiv:2610.15634v1cs.AI

TL;DR

Evaluative judgements often lack a single verifiable answer, complicating the assessment of sycophancy. Big-Big-SycoBench tests score revisions after valid and invalid arguments against each model’s baseline variability. Across ten models, responsiveness to valid arguments is often high, while compliance with invalid arguments varies widely.

  • Problem

    Evaluative tasks can have several defensible conclusions rather than a single verifiable ground truth, making sycophantic agreement harder to detect.

    Source: §1 · p. 2

  • Method

    Big-Big-SycoBench measures score revisions after valid or invalid arguments, counting updates only when they exceed twice the model’s baseline standard deviation.

    Source: §1 · p. 3, §3.3 · p. 6

  • Results

    Seven of ten models follow 88–95% of valid arguments, while compliance with invalid arguments ranges from 0.15% to 94.6%.

    Source: §5 · p. 8

  • Takeaways & Limitations

    GPT-6 Astra follows 88.3% of valid arguments and 0.15% of invalid ones, showing that low compliance can coexist with high responsiveness.

    Source: §5 · p. 8

    The benchmark measures numerical score changes after a single challenge, not the correctness of ratings or whether their explanations support revisions.

    Source: §5.1 · p. 9

Abstract

from arXiv · show

Sycophancy in large language models risks amplifying unsupported claims or beliefs across the knowledge ecosystem as these systems increasingly contribute to research synthesis, evaluation, and decision-making. A reliable model should revise its judgement when a user identifies a genuine flaw, but should resist pushback based on fallacies, emotion, or irrelevant appeals to authority. While sycophancy is usually measured on questions with a verifiable answer, the evaluative tasks in which it is most difficult to detect have no such ground truth. We introduce Big-Big-SycoBench, which measures sycophancy through argument discernment: the ability to revise an evaluation after valid criticism while resisting invalid pressure. The benchmark consists of 450 text artefacts with 5,400 arguments for raising or lowering their quality scores, each reviewed by one or more of twelve human annotators. An LLM scores each artefact twenty times to establish its baseline score and uncertainty. The model is then given a valid or invalid argument and asked to update its score. We only count a revision if it exceeds the model's own baseline variability. We use the resulting score delta to distinguish responsiveness to valid arguments from compliance with invalid ones. Across ten models, responsiveness is 88-95\% for seven, 76-77\% for Kimi K3 and GLM-5.3, and 32\% for Opus 5.5. Compliance ranges from 0.2\% for GPT-6 Astra to 95\% for GPT-4o. Big-Big-SycoBench is intended to guide the development of models whose judgements can be corrected by sound arguments, but not displaced by unsupported pressure.

1 Introduction

Big-Big-SycoBench tests whether models revise evaluative judgements for sound reasons, rather than simply yielding to user pressure. Across ten models, responsiveness to valid arguments is often high, but compliance with invalid arguments varies widely.

1 Introduction

  • Prior work links agreeable post-training preferences with validating incorrect beliefs, while other benchmarks examine preference reproduction, sustained pressure, or misleading suggestions.The introduction positions this benchmark around argument merit in evaluative judgements without a single ground-truth answer.Source: §1 · p. 2
  • Because models contribute to research synthesis, peer review, and scientific production, yielding to pressure in these roles can amplify unsupported claims across the knowledge ecosystem.The cited examples include literature screening, manuscript feedback, and scientific production.Source: §1 · p. 2
  • User feedback is not inherently sycophantic: models should revise after genuine criticism while resisting unsupported assertions, irrelevant authority appeals, and emotional pressure.The benchmark also records stubbornness and contrarian movement, not only compliance.Source: §1 · pp. 2–3
  • Evaluative tasks lack a single verifiable ground truth, making sycophantic agreement harder to distinguish from independent judgement.Such tasks include reviewing proposals and assessing policy memos, where several conclusions may be defensible.Source: §1 · pp. 2–3
  • The benchmark uses 450 artefacts and 5,400 human-reviewed arguments to test whether models revise scores according to argument merit without requiring a single correct score.Arguments either identify genuine strengths or weaknesses or advocate a change without substantive support.Source: §1 · p. 3
  • Seven of ten models follow 88–95% of valid arguments, while invalid-argument compliance ranges from 0.2% for GPT-6 Astra to 95% for GPT-4o.The evaluation covers five frontier and five legacy OpenAI-family models alongside models from other developers.Source: Figure 1 · p. 2, §1 · p. 3

2 Benchmark construction

Big-Big-SycoBench controls artefact properties and argument validity through generation, human review, revision, and verification. Its released data balance domains, target qualities, revision directions, and validity conditions.

2 Benchmark construction

  • The benchmark pairs 450 artefacts with twelve arguments each, distinguishing arguments grounded in genuine strengths or weaknesses from unsupported attempts to change scores.The resulting 5,400 arguments cover raising and lowering quality scores.Source: §2 · p. 3

2.1 Stage 1: controlled data generation

  • Specifications controlled artefact length, domain, target quality, and charged framing; each retained artefact had planted strengths and weaknesses for arguments in both revision directions.Each artefact received twelve distinct arguments across a 2 × 2 × 3 direction–validity design, matched in persuasive register.Source: §2.1 · p. 4

2.2 Stage 2: human review

  • Twelve annotators reviewed all artefacts and associated arguments, with deliberate overlap; pairwise agreement on label acceptability across overlapping reviews was 94.1%.Review notes informed revision rather than serving as a binary inclusion rule.Source: §2.2 · p. 4

2.3 Stage 3: revision and argument normalisation

  • All 5,400 arguments were revised or normalised using QC and review evidence to remove construction cues while preserving intended direction, validity, register, and approximate length.Automated pattern checks screened for residual cues after normalisation.Source: §2.3 · p. 5

2.4 Stage 4: final verification

  • Final verification used an automated classifier and human adjudication of uncertain or undecided cases, producing 133 targeted repairs across 97 artefacts.No unresolved verification cases remained in the final dataset.Source: §2.4 · p. 5

2.5 Stage 5: final dataset

  • The release balances ten domains at 45 artefacts each and the four direction–validity combinations at 1,350 arguments each; generator anchor scores are not ground-truth ratings.Evaluated models do not receive validity labels, quality rationales, planted properties, or review metadata.Source: §2.5 · p. 5

3 Benchmark evaluation

Evaluation establishes each model’s baseline score variability, then measures its response to valid and invalid arguments. The benchmark reports separate responsiveness and compliance scores, with updates counted only when they exceed baseline variability.

3.1 Baseline scoring

  • Each artefact is scored across 20 independent runs, and the baseline mean and standard deviation capture the model’s score distribution without assuming a single ground-truth rating.The benchmark measures subsequent changes relative to this baseline.Source: §3.1 · p. 5

3.2 Rescoring in response to valid/invalid arguments

  • Each baseline conversation is extended with an argument in the direction opposite the baseline score’s side of 50, reducing ceiling effects.Three valid and three invalid arguments are used for raising or lowering scores.Source: §3.2 · p. 6

3.3 Computing weighted update probabilities

  • Following and contrarian updates count only when score change exceeds twice the baseline standard deviation; changes within that band count as stable.Argument-quality weights estimate how clearly each argument earns its validity label before rates are averaged across artefacts.Source: §3.3 · p. 6

3.4 The final benchmark score

  • The final benchmark reports follow-minus-contrarian probabilities separately for valid arguments (responsiveness) and invalid arguments (compliance), reflecting sycophancy’s multiple dimensions.The leaderboard ranks models by increasing compliance.Source: §3.4 · p. 6

4 Results

Across the evaluated models, valid-argument following is high for most, but invalid-argument compliance varies widely. Valid-following rates also differ by challenge direction, and Opus 5.5 shows especially low responsiveness.

4 Results

  • Nine models were evaluated on all 450 artefacts; Opus 5.5 used a stratified 100-artefact subset to limit its estimated $3,700 full-benchmark cost.The subset preserves domain and length proportions and was selected without using model scores.Source: §4 · p. 6

4.1 Compliance varies widely despite high responsiveness

  • Invalid-argument following ranges from 0.15% for GPT-6 Astra to 94.6% for GPT-4o, while valid following is 88.3–95.3% for six OpenAI models and Grok 4.7.Kimi K3 and GLM-5.3 have valid-following rates of 76.2% and 77.4%, respectively.Source: §4.1 · pp. 6–7, Figure 4 · p. 7
  • GPT-4o follows invalid arguments more often than valid ones (94.6% versus 89.0%), and both GPT-5.6 variants follow invalid arguments more often than GPT-5.5.Contrarian movement remains below 0.2% for every model and feedback class.Source: §4.1 · p. 7

4.2 Valid following differs between lowering and raising arguments

  • Across all nine models tested in both directions, valid-following rates are higher for lowering than raising arguments; GPT-6 Astra is 97.6% versus 63.9%, and GLM-5.3 is 84.7% versus 6.3%.These comparisons use different artefact groups, selected by each model’s mean baseline score; the observed differences may also reflect artefact characteristics and selected arguments.Source: Figure 5 · p. 7, §4.2 · pp. 7–8

4.3 Opus combines low compliance with lower responsiveness

  • On its 100-artefact subset, Opus 5.5 follows 0.8% of invalid and 32.2% of valid arguments, while other models’ valid-following rates on that subset range from 75.1–95.0%.At a zero threshold, Opus follows 78% of valid lowering but only 1% of valid raising arguments; the authors suggest its near-absence of upward revision reflects reluctance to raise scores.Source: §4.3 · p. 8

5 Discussion and Conclusion

Big-Big-SycoBench separates revision after valid arguments from compliance with invalid ones, revealing distinct model patterns. Its single-challenge numerical evaluations do not establish rating correctness or whether stability after valid criticism is justified.

5 Discussion and Conclusion

  • Seven of ten models revise after 88–95% of valid arguments, while compliance spans 0.15% to 94.6%; GPT-6 Astra follows 88.3% of valid arguments and 0.15% of invalid ones.GPT-4o follows invalid arguments slightly more often than valid ones; contrarian movement is essentially absent.Source: §5 · p. 8
  • On the shared 100-artefact subset, Opus 5.5 follows 32.2% of valid arguments, versus 75.1–95.0% for the other models, despite ranking second by compliance.Kimi K3 and GLM-5.3 combine compliance below 14% with responsiveness of 76–77%.Source: §5 · p. 8
  • Within the OpenAI family, compliance falls from 94.6% for GPT-4o to 0.15% for GPT-6 Astra while responsiveness stays within 88–95%.The decline is not monotonic, and models of similar capability differ in compliance by more than a factor of two; the design does not identify responsible training choices or reasoning budgets.Source: §5 · p. 8
  • Every model tested in both directions follows valid lowering arguments more often than valid raising arguments, though the raising group consists of artefacts rated below 50.For manuscript- or proposal-assessing models, this suggests valid author defences of poorly rated texts may be the least likely to prompt revision.Source: §5 · p. 8
  • Invalid arguments use appeals to authority and consensus, emotional appeals, and bare assertions; the two lowest-compliance models resist these cues in more than 99% of cases.GPT-4o followed these invalid cues as readily as valid criticism.Source: §5 · p. 8

5.1 Limitations and future work

  • The benchmark measures numerical evaluation changes after a single challenge, not rating correctness, explanation quality, or whether stability after a valid argument is justified.Baseline standard deviation measures repeatability, not necessarily epistemic uncertainty; the update threshold is a practical rule, and models receive different arguments for the same artefact.Source: §5.1 · p. 9
  • Generalisation to naturally occurring criticism, other languages, and real evaluation workflows remains to be tested.Proposed future work includes neutral rescoring controls, shared artefacts tested in both directions, human judgements of revisions, independent weight judges, and human rescoring.Source: §5.1 · p. 9

AI use statement

Generative AI assistants supported multiple stages of the work, and language models also generated and evaluated benchmark materials. The authors reviewed and verified the work and take responsibility for its final content.

AI use statement

  • Generative AI assistants assisted with literature retrieval, citation verification, language revision, drafting, organisation, design feedback, implementation, interpretation, and consistency checks.Source: AI use statement · p. 9
  • Language models generated the artefacts and challenge arguments, reviewed validity labels, revised flagged arguments, and produced pairwise argument-quality judgements for Bradley–Terry weights.Source: AI use statement · p. 9
  • The authors reviewed and verified the text, code, references, analyses, and claims, and take responsibility for the final work.Source: AI use statement · p. 9

Ethics statement

The benchmark uses synthetic, model-generated artefacts and arguments, with annotators drawn from the authors’ research lab. Its results characterize responses under pressure but do not certify suitability for real-work review or grading.

Ethics statement

  • All benchmark artefacts and challenge arguments are synthetic, model-generated texts, and the study uses no human subjects or personal data.Source: Ethics statement · p. 9
  • The twelve annotators are research-lab members and paper co-authors who assessed validity labels for synthetic texts.Source: Ethics statement · p. 9
  • Invalid arguments are common rhetorical fallacies rather than adversarial attack material, and charged topics use good-faith framings.Six artefacts with fabricated or pseudo-real sources were edited to mark those references as fictional.Source: Ethics statement · p. 9
  • The findings characterize model revisions under pressure but do not certify suitability for reviewing or grading real work.Source: Ethics statement · p. 9

Reproducibility statement

The reproducibility statement describes planned public release of code and data, alongside a released unlabeled subsample for exploration. It points to sections and figures covering construction, scoring, diagnostics, and release details.

Reproducibility statement

  • The authors state that code and data will be released in public repositories upon acceptance.Source: Reproducibility statement · p. 9
  • A subsample of 100 artefacts and 1,200 arguments without validity labels has been released on Hugging Face for data exploration.Source: Reproducibility statement · p. 9
  • The paper directs readers to sections and figures for benchmark construction, evaluation, scoring, bootstrap intervals, alternative scores, diagnostics, threshold sensitivity, dataset release, and software dependencies.Source: Reproducibility statement · p. 9

A.1 Model specifications

The evaluation used ten models accessed through official endpoints, with model-specific APIs and reasoning settings. All models except Opus 5.5 were evaluated on 450 artefacts; Opus 5.5 used a stratified subset of 100.

Model specifications

  • All six OpenAI models used the Responses API through batch processing, while Opus 5.5 used Anthropic’s Messages Batch API and the other listed models used OpenRouter.Sampling parameters were left at provider defaults; all models used the maximum token allowance, with hard-coded limits where necessary.Source: §A.1 · p. 14

A.2 Benchmark construction details

The appendices detail how the benchmark data and model-update measures were constructed, how argument quality was weighted, and how alternative aggregate scores and contrarian responses are treated.

A.2 Benchmark construction details

  • The 450-artefact dataset spans ten domains, three target lengths, and five target quality levels; one quarter of artefacts in four eligible domains used charged but good-faith framing, producing 45.Flagged arguments were revised to preserve intended direction and validity condition. After revision, regular-expression checks screened for residual construction cues.Source: §A.2 · pp. 14–15

A.3 Model updates

  • The update protocol compares each new score with the initial score and counts shifts beyond δi = 2σ0i as updates; σ0i measures scoring variability on that artefact.A negative shift indicates a response opposite to the argument’s push.Source: §A.3 · p. 15, §A.4 · p. 16

A.4 Benchmark score

  • Argument-quality weighting produces separate following, stable, and contrarian rates for valid and invalid feedback.Arguments are weighted by how clearly they earn their validity label; arguments rated on the wrong side of the boundary carry zero weight.Source: §A.4 · p. 16

A.5 Argument quality score

  • Each argument’s Bradley–Terry score represents its quality, with higher scores indicating greater validity and greater merit for an update.Argument weights encode label confidence; arguments on the wrong side of the validity boundary receive zero weight, and clearly invalid arguments are penalized more strongly than borderline ones.Source: §A.5 · pp. 16–17

A.6 Alternative final scores

  • Alternative aggregate scores can conceal distinct behaviours: models with different valid-following and invalid-following rates may receive the same difference score.The authors therefore favor the six update probabilities, or their visual representation, as a more complete behavioural picture.Source: §A.6 · p. 17, Figure A1 · p. 18

A.7 Contrarian responses

  • Contrarian response rates are at most 0.012% for valid arguments and 0.189% for invalid arguments across the ten models at the reference threshold.The two following probabilities approximate the six-probability profile in these results, but small averages do not rule out larger rates on individual arguments.Source: §A.7 · p. 18

A.8 Evaluation cost

Evaluation costs varied by model, and the cost of repeated scoring led the authors to evaluate Opus 5.5 on a 100-artefact subset.

A.8 Evaluation cost

  • Most frontier models cost $200–$600 per run; GPT-6 Astra and Kimi K3 cost approximately $1,000, while GPT-5.6 Luna cost under $40.The Opus 5.5 subset cost below $800, versus an estimated $3,700 for a full evaluation.Source: §A.8 · p. 19

A.9 Data and software

The appendix records the data and software used and provides supplementary figures on baseline scores, threshold sensitivity, update magnitude, and score variability.

A.9 Data and software

  • The dataset and accompanying code are slated for release in public repositories on acceptance; model identifiers, prompts, and request settings are recorded in run configurations.The experiments used Python 3.14.2 and listed analysis, visualization, Bradley–Terry, and API software packages.Source: §A.9 · p. 19

B Additional tables and figures

  • Figures B1 and B2 summarize baseline score and variability distributions and initial scores by target argument quality; Opus 5.5 uses 100 artefacts while other models use 450.The B2 anchor values are construction targets, not observed-score cutoffs.Source: Figure B1 · p. 20, Figure B2 · p. 21
  • Figures B3 and B4 show sensitivity of following probabilities to alternative thresholds; B3 compares threshold rules, while B4 varies k around the default k = 2.These are scoring sensitivity checks, not calibrated significance tests.Source: Figure B3 · p. 22, Figure B4 · p. 22
  • Figure B5 plots magnitude-aware updates for invalid and valid feedback, capturing both update frequency and movement beyond the stability threshold.The caption cautions that larger valid updates are not necessarily better and that no confidence intervals are shown.Source: Figure B5 · p. 23
  • Figure B6 compares valid and invalid score revisions across artefact-variability bins; models differ in whether revisions track variability and argument validity.Astra’s invalid revisions remain near zero, while GPT-5 and GPT-4o revise more after invalid than valid arguments; Opus 5.5’s valid revisions remain low without clear dependence on variability.Source: Figure B6 · p. 24

Related papers

Loading 2610.15634v1…