acceptodds
Under review as a conference paper at ICLR 2027

Design, Not Noise: Certifying Cross-Model Claims in Black-Box Behavioural Audits of Language Models

Abstract

Behavioural benchmarks (economic games, sycophancy, dark patterns) report a leaderboard: one number per model, under one elicitation design chosen by the authors. We show that the published record cannot adjudicate its own leaderboards. Across 28 fully crossed modeldesign tables from the six of eight published behavioural benchmarks that print one, more than one model attains rank 1 in 70.6% of the 17 tables whose crossed facet the benchmark itself calls irrelevant, and 25.8% of their 612 pairwise orderings reverse on one of the papers' own variants, rising to 55.0% of the 60 that vary a prompt paraphrase (all from GAMA-Bench, the only benchmark publishing one). No crossed table publishes repetitions, so no reversal can be attributed to design rather than noise. We supply the missing instrument: a design-relative estimand with an identified interval, a three-level certificate hierarchy with valid tests, a closed-form breakdown tilt, and a replicate floor that triages every claim as certified, noise-limited or design-limited. Run over 9 models 4 domains 96 crossed designs, each domain yields 73–94 distinct full leaderboards and uniform dominance certifies 0–8% of pairs; seven 2026-frontier models are no more stable (8 of 84 pairs certified) and show no detectable capability attenuation (, model-clustered se ). Decidability is an item budget: taking the worst-served domain from 16 items to 96 drops its noise floor while the design signal holds, and its undecidable claims from 26 of 36 to 10. We release the corpus, the toolkit and a reporting standard: a point estimate and an identified interval per model, a certificate per claim. All code, data and raw responses accompany this submission as supplementary material and will be released at https://github.com/anonymous-iclr2027/design-relative-behaviour upon acceptance.

Primary area:
datasets and benchmarks
Keywords:
behavioural benchmarks, evaluation design, elicitation sensitivity, simultaneous inference, model ranking, auditing, llm evaluation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_SCxPOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
3
Presentation
2
Soundness
3(+1)
Read review
Paper summary

The paper studies whether cross-model behavioural comparisons remain defensible when elicitation designs change. It audits published benchmarks and introduces a design-relative framework distinguishing uniform dominance, mean ordering and single-design claims. The framework includes nuisance averaging, framing-dependent intervals, a breakdown-tilt measure, and a procedure classifying comparisons as certified, noise-limited, or design-limited. Experiments cross models, tasks, elicitation designs, and repeated responses, with additional item-budget, judge, and frontier-model analyses. The central contribution is a framework intended to distinguish genuine design sensitivity from instability caused by finite sampling.

Critical strengths
  1. The paper usefully separates claims and estimands. Distinguishing single-design performance, a specified average, and dominance across all enumerated designs makes explicit what a leaderboard comparison does and does not establish. The breakdown-tilt measure provides an interpretable sensitivity summary instead of another aggregate robustness score.

  2. The experiments make a substantial effort to distinguish sensitivity from measurement noise. The crossed design, repeated observations, item-splitting analysis and larger-item follow-up go well beyond merely demonstrating prompt-induced rank changes. The paper also appropriately acknowledges that the published record often lacks enough information to attribute reversals to design.

Critical weaknesses
  1. The 'designed-limits' verdict is not sufficiently established by the stated tests (Section 4.3 and Appendices B-C). Rejection of the permutation null provides evidence against design exchangeability, but does not by itself establish that the true pairwise contrast changes sign across designs. A pair could have mostly very small positive contrasts and a few large positive contrasts: design dependence may be detectable, noisy estimates may reverse near zero, and the overall mean gap may exceed 0.04, though the true ordering remains uniform. The paper itself notes that DIS and the split-sample floor are different functionals and that floor conservativeness requires per-design separartion. Restricting the average gap doe not supply that condition. Consequently, the inference that additional sampling cannot resolve a comparison labelled design-limited needs stronger justification. Calibration under a design-invariant null does not settle the heterogenous, same-sign case.

  2. The mapping from published claims to uniform dominance is sometimes too strong (Section 4.2 and Appendix F). For example, a statement that some models favor their developers' products does not inherently assert strict pairwise dominance under every permissible elicitation. The conclusion that seven of eight audited papers require L1 therefore depends on an interpretation that needs more explicit justification. This matters because the paper's criticism of existing conclusions is stronger than its demonstration of sensitivity: showing that a result changes across designs does not automatically establish that the original paper claimed invariance across those designs.

Questions for authors
  1. How does the full triage procedure behave under heterogeneous but strictly same-sign true contrasts, including many near-zero contrasts? In particular, what is the false 'design-limited' classification rate in that setting? Can the existing data support both a positive and a negative true contrast, with appropriate uncertainty assessment, for the comparisons labelled design-limited?

  2. Can you distinguish genuinely design-uniform claims in the audited papers from average, prevalence, or explicitly conditional claims, and explain how this distinction affects the seven-of-eight conclusion?

Reviewer_XRsCOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper shows that behavioral benchmarks for language models usually use a single prompt. Prompt changes that do not affect the task can change the model rankings. Without repeated tests, it is not possible to know if these changes are real effects or sampling noise. The paper proposes a framework that calculates averages across irrelevant prompt variations, reports result ranges for meaningful prompt variations, assigns a strength grade to each ranking claim, and uses a noise floor to distinguish design effects from sampling noise. A large, fully crossed experiment shows that few ranking claims are reliable. The paper also proposes a new standard for reporting benchmark results.

Critical strengths

The paper changes a known concern about prompt sensitivity into a practical, low-cost procedure. Benchmark authors can use this procedure to specify the strength of each ranking claim. They can also identify whether design choices or sampling noise cause unstable results. This procedure makes behavioral benchmarking more transparent and fairer than a leaderboard that uses one prompt. However, it does not make benchmarking fully fair. The selected prompt variations, answer formats, and evaluation settings continue to limit the results.

Critical weaknesses
  1. It is well known that prompt wording can affect LLM outputs and rankings. It is also common practice to evaluate a model with different prompt wordings. Thus, the main contribution is a formal method for recording and reporting results, not a new finding.
  2. The main experiment uses mostly forced-choice answers or single-number answers. It also uses very short completion limits. Thus, the experiment measures which option a model selects. It does not fully measure how a model behaves when it explains its answer or gives its reasoning. The paper cites evidence that constrained and open-ended prompts can produce opposite conclusions. The open-response part of the experiment is small, and its results are mostly inconclusive. Therefore, the findings might not apply to evaluations of complete responses.
Questions for authors

Since your main results rely on short forced-choice answers, what evidence suggests they would hold for full, explained responses?

Reviewer_2YwTOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

Paper appears to argue that there is systemic mismatch in reported scores on what the authors call behavioral benchmarks and what can be reproduced. The paper attempts to remedy this with CIs around the scores.

Critical strengths

I agree that benchmarks on these broad domains such as "economic games, sycophancy, dark patterns" are often unreliable and do not hold up upon attempted replication. Work to remedy this issue is valuable.

Critical weaknesses

The paper is extremely difficult to parse and the results are uninterpretable. When the crux of the paper is to better elucidate the performance of models on important benchmarks, I believe clear presentation of the results is paramount.

Questions for authors

None at this time.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.