Pith. sign in

REVIEW 3 major objections 6 minor 65 references

What Current AI Benchmarks Leave Unmeasured: Modality, Search, Citations, and Implications (for Safety Evaluations)

T0 review · 3 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Even within one model family, benchmark scores shift with access interface and web search; accuracy alone hides inconsistency in answers, citations, and refusals.

desk verdict A carefully run audit showing deployed chat behavior differs from API behavior in ways accuracy metrics miss; the 'modality' label overreaches slightly because model weights are unverified. read the letter →

arxiv 2608.06202 v1 pith:2LU6EFPN submitted 2026-08-06 cs.HC cs.AI

Ro Encarnación , Tina Behzad , Emma Lurie , Danaé Metaxa This is my paper
classification cs.HCcs.AI
keywords LLMbenchmarkingAIsafetyevaluationaccessmodalitywebsearchresponseconsistencycitationgroundingabstentionbehaviorbenchmarkaudit
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper argues that the way LLM benchmark results are obtained—through a single access interface (an API), one run per prompt, and accuracy as the headline number—can conceal behavior that matters for AI safety. To test this, the authors compared the chat interface and API of a widely used model, with and without web search, across 401 prompts from the BBQ and SafetyBench benchmarks, running each prompt three times per condition. They found that accuracy alone missed substantial variation: repeated runs disagreed up to 21% of the time, web search reduced accuracy by up to 8 percentage points and reversed the accuracy gap between interfaces on SafetyBench, and the two interfaces cited almost entirely different sources. The conclusion is that reporting only accuracy from a single API-based run overstates the reliability of deployed systems, and that safety evaluations should treat interface, search condition, multi-run consistency, citation grounding, and abstention as standard measured dimensions.

What carries the argument

The central object is a 2×2 factorial audit design crossing modality (chat UI vs API) with web search (enabled vs disabled), repeating each prompt three times per condition to produce within-modality and between-modality response pairs. The methodological engine is the comparison of between-modality disagreement and similarity against within-modality run-to-run variability, modeled with generalized and linear mixed-effects models that include prompt-level random intercepts, so that systematic interface effects can be separated from ordinary model stochasticity. A second element is the expansion of outcome measurement beyond accuracy to response consistency, lexical and semantic text similarity, citation grounding, and abstention behavior.

What would settle it

Obtain exact serving checkpoint identifiers (hashes or manifests) for the API model gpt-5.3-chat-latest and the logged-out chat interface model during the audit window: if the checkpoints differ, the modality contrast is confounded with model identity and the paper's central interpretation falls. A complementary test is to rerun this 2×2 audit on a model family where API and chat checkpoints are verifyably identical and check whether the consistency, citation, and abstention gaps persist.

Watch

Extended reading notes

Core claim

The paper reports a 2×2 audit of one widely used LLM family in which the same 401 prompts from the BBQ and SafetyBench benchmarks were run through the chat interface and the API, with and without web search, three times each. The authors find that accuracy differences between interfaces and search conditions, though only a few percentage points, were statistically significant, reversed direction on SafetyBench when search was enabled, and were accompanied by larger differences in dimensions standard evaluations do not report: repeated runs of the same prompt picked different answers up to 21% of the time; API and chat UI responses were lexically and semantically more different from each other than from their own repeated runs; the two interfaces cited overlapping sources only about 4% of the time at the URL level; and abstentions were rare but unreproducible, with the same prompt refusing in one run and answering in the next. The paper's conclusion is that accuracy-focused, API-only, single-run benchmark reporting overstates the reliability of deployed systems and that AI safety evaluations should treat modality, search, multi-run consistency, citation grounding, and abstention as standard measured dimensions.

Load-bearing premise

The load-bearing premise is that the API model gpt-5.3-chat-latest and the GPT-5.3 Instant served in the logged-out chat interface are the same underlying model; the audit relies on the provider's public documentation for this, so if the chat interface serves a differently tuned or distilled checkpoint, the measured modality differences could actually be model differences.

Editorial extensions

If this is right

  • Safety evaluations that report only API-based accuracy can miss the interface most users actually encounter, so a model that scores well in the benchmark may behave differently in the chat UI.
  • Enabling web search is not neutral: it reduced accuracy in both modalities and increased answer inconsistency on both benchmarks, so evaluations should specify search conditions rather than assume search leaves behavior unchanged.
  • Because repeated runs of the same prompt disagreed up to 21% of the time, single-run evaluation cannot support claims about model reliability; multi-run consistency reporting should become standard.
  • API and chat UI ground answers in almost disjoint sources even when the answers agree, so citation-grounding results from one access layer cannot be extrapolated to the other.
  • Abstention inconsistency means resubmitting a prompt can bypass a refusal, so accuracy-only scores treat a guardrail that fails on retry as equivalent to one that never refuses.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If these results generalize to other providers and model versions, benchmark leaderboards are best read as point estimates for one access configuration, and any regulation relying on a single pass rate should require multi-condition, multi-run reporting.
  • A deployment-weighted follow-up could quantify the practical severity: since most consumer traffic goes through the chat UI, weighting accuracy by actual usage could make API-only scores misleading in a direction current benchmarks do not capture.
  • The near-zero citation overlap suggests the two interfaces may draw on different retrieval backends or source-selection policies; auditing the retrieved pools directly, rather than only the cited text, would isolate whether the difference lies in search, summarization, or presentation.
  • The refusal inconsistency implies safety guardrails should be stress-tested by resubmission across runs and interfaces, since a guardrail that answers on retry provides no reliable protection.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper audits whether benchmark outcomes for two safety/bias benchmarks (BBQ and SafetyBench) are stable across access modality (ChatGPT chat UI vs OpenAI API) and web search condition. It samples 401 prompts, runs each three times per condition (4,812 responses total), and measures accuracy, response consistency, Jaccard and cosine similarity, citation grounding, and abstention behavior. The paper reports small but significant accuracy differences, a SafetyBench accuracy reversal under search, higher between-modality than within-modality answer disagreement (OR = 1.51), low citation overlap between modalities, and rare, run-varying abstentions. It argues that standard API-based, single-run, accuracy-only benchmark evaluations miss safety-relevant behavioral variation in deployed systems.

Significance. If the results stand, the paper makes a useful empirical contribution to the evaluation-methodology literature. Its design is transparent: an a priori power analysis, stratified sampling, mixed-effects models with prompt-level random intercepts, Benjamini-Hochberg correction for primary tests, and a public repository for datasets and code. It also measures multiple behavioral dimensions beyond accuracy, which is a genuine strength. The central limitation is that the modality contrast assumes identical model weights across the chat UI and the API, and this assumption is not verified or explicitly acknowledged. That confound weakens the causal reading of the between-modality results, although the within-modality search effects and the multi-run inconsistency findings are less affected by it and remain valuable.

major comments (3)
  1. [3.1 (Model and environment selection)] Section 3.1 states that the logged-out chat UI serves GPT-5.3 Instant and that the API audit uses gpt-5.3-chat-latest 'for comparability,' but this identity is unverified. OpenAI does not publicly expose the weights served to the chat interface, and the chat-deployed model could be distilled, safety-tuned, or otherwise differently configured. Because the paper's definition of modality explicitly includes system prompts, search, and moderation but not model weights, every between-modality result (the SafetyBench accuracy reversal in Table 1, the disagreement OR of 1.51 in Section 4.2, the similarity gaps in Section 4.3, and the citation overlap in Section 4.4) could in principle be driven by model variant rather than access layer. The phrase 'even within a single model family' is too strong under this uncertainty. Please either provide behavioral evidence that the two endpoints are the same underlying model (e.g., matched behavioral or logit fingerprints on a probe set) or reframe the central claim as a comparison of deployed access conditions, explicitly stating that model weights are a potential confound.
  2. [4.4 (Citation rate and source overlap)] The citation analyses mix two different source types. Section 3.2 says chat UI citations include both in-text URLs and supplementary sources in the 'More' side panel, while the API returns only in-text citations. Section 4.4 then reports that the chat UI has a higher citation rate and a larger unique-source set (n = 356 vs 179 on BBQ) but switches to 'restricting to the chat UI's in-text citations' for the overlap analysis. The headline gap in citation rate and unique-source counts may therefore be an artifact of an interface feature that the API cannot access. Please report citation rate and unique-source counts separately for in-text and supplementary chat citations, and recompute the per-prompt union and overlap with a single consistent source definition.
  3. [4.5 and 5.1 (Abstention evidence)] The abstention analysis rests on six events across four BBQ prompts, with no SafetyBench abstentions. The descriptive claim that abstentions were inconsistent across runs is supported (three of four prompts abstained on only one of three runs). However, the Discussion's conclusion that 'if a model guardrail can be bypassed in any interface by simply resubmitting a question and receiving a different output, then the guardrail is functionally useless' is disproportionate to this evidence base. Please either substantially soften the safety conclusion to a hypothesis for future research or expand the abstention sample beyond four prompt IDs.
minor comments (6)
  1. [3.2 / 4.4] Clarify whether the supplementary 'More' citations are included in the reported citation-rate percentages and unique-source counts; the text appears to switch definitions between the two analyses.
  2. [4.2] Specify that the consistency GLMM outcome is binary 'all three runs agree,' so an odds ratio of 0.51 for search indicates lower odds of consistency (i.e., more inconsistency); the current wording could be misread as search reducing inconsistency.
  3. [4.5] The statement that 'the same prompt refused in one setting was often answered in another' overstates the evidence; only one of the four abstaining prompts abstained in both modalities, so 'occasionally' is more accurate.
  4. [A.2 vs Table 1 footnote] The Appendix says 7 of 2,376 BBQ runs could not be parsed, while the Table 1 footnote says one BBQ chat no-search response was excluded; align these counts and explain the discrepancy.
  5. [Figure 3 and Section 4.3] The example text appears to mix the API/no-search and API/search runs: the API response in Figure 3 contains a citation, while the surrounding text says API/no-search responses contained only the answer choice. Clarify which condition is shown.
  6. [A.4 (Chat interface pipeline)] In the no-search chat condition, 'no configuration step was performed' rather than explicitly disabling search; specify how the absence of search was verified (e.g., no citations, no search tool calls) to ensure the condition is well-defined.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the audit reports external measurements against fixed benchmark gold labels, and its claims do not reduce to fitted inputs or self-citation.

full rationale

This paper is an observational audit, not a derivation or a prediction from fitted parameters. Accuracy is scored against the fixed gold labels of two external benchmarks (BBQ and SafetyBench), and all other measures—consistency, Jaccard/cosine similarity, citation overlap, abstention—are computed directly from collected responses. No quantity is fit to a subset of data and then presented as a prediction of a closely related quantity, and no target result is encoded in the definition of the independent variables. The central contrast between chat UI and API is defined in terms of the full user-facing system, but the paper states this definition explicitly and treats it as the intervention under audit rather than as a hidden equivalence; the finding that outcomes differ is then an empirical result, not a tautology. The paper's invocation of prior work, including a citation that shares a co-author (Proebsting et al. 2025), is used only to echo prior refusal-rate findings and is not load-bearing for the paper's own conclusions. The main validity threat—that the logged-out chat UI may serve different weights than the API model gpt-5.3-chat-latest—concerns external validity and model identity, not circularity: even if the weights differ, the paper would still be reporting real differences between two deployed access conditions, and its statistics would remain as computed. No self-referential definition, fitted-parameter-as-prediction, or self-citation chain forces the reported results.

Assumptions & free parameters 0 free parameters · 5 assumptions · 0 invented entities

The audit is empirical, so the ledger is lean. The main epistemic burden is the model identity assumption (second axiom) and the validity of benchmark labels (first axiom). The statistical models estimate effects rather than introducing hidden free parameters.

assumptions (5)
  • domain assumption BBQ and SafetyBench gold labels validly measure social bias and safety.
    Used throughout; accuracy scoring compares model selections to these gold labels (Section 3.2).
  • domain assumption API endpoint gpt-5.3-chat-latest corresponds to the same model as the chat UI default GPT-5.3 Instant.
    Section 3.1; needed for modality comparisons to isolate the access layer.
  • domain assumption Logged-out chat UI sessions without personalization represent a meaningful deployed baseline.
    Section 3.1; the audit intentionally avoids account personalization, which may differ from typical logged-in usage.
  • domain assumption Exact-string matching with manual review correctly identifies the model's selected answer choice.
    Section A.2; unresolvable responses were handled via manual review (BBQ) or random assignment following benchmark protocol (SafetyBench).
  • domain assumption Repeated runs are independent draws from the model's stochastic decoding.
    Section 3.1; browser sessions and API calls are isolated, supporting consistency measures.

how reviews work

0 comments
Cite this review

Pith. "Pith review of What Current AI Benchmarks Leave Unmeasured: Modality, Search, Citations, and Implications (for Safety Evaluations)." pith.science (2026). https://pith.science/paper/2LU6EFPN

@misc{pith2026260806202,
  author       = {Pith},
  title        = {Pith review of: What Current AI Benchmarks Leave Unmeasured: Modality, Search, Citations, and Implications (for Safety Evaluations)},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2LU6EFPN}},
  note         = {Machine review of arXiv:2608.06202}
}
read the original abstract

Large language model (LLM) benchmark evaluations are routinely used to support claims about model safety, reliability, and deployment readiness. Yet most evaluations rely on a single access modality (model APIs), perform a single run per prompt, and report accuracy as the primary outcome metric, without accounting for conditions such as web search that may have effects on model behavior in deployment. We audit these assumptions for one of the most widely-used LLMs, comparing two modalities, ChatGPT's chat UI and OpenAI's API, with and without web search enabled. We use a stratified total sample of 401 prompts from two popular benchmarks, BBQ and SafetyBench, collecting 4,812 total responses across three repeated runs per prompt. Beyond standard performance measures, we evaluate model output dimensions including response consistency, response text similarity, citation grounding, and abstention behavior. For instance, chat UI responses were less accurate than API responses on both benchmarks with search disabled. Enabling web search reduced accuracy by up to 8 percentage points, and even reversed the direction of modality performance trends for one benchmark. Repeated runs of the same prompt produced inconsistent responses in up to 21\% of prompts. The two modalities also grounded answers in different citations, and abstention behavior was also inconsistent across both modalities. These results illustrate that, even within a model family, reporting only simple accuracy metrics can obscure important forms of model behavioral variation relevant to AI safety assessments. We argue that AI safety evaluations should systematically account for modality, multi-run consistency, search conditions, and response-level behaviors to better reflect how deployed AI systems behave in practice.

Figures

Figures reproduced from arXiv: 2608.06202 by the authors.

Figure 1
Figure 1. Average per-run accuracy across modalities and search conditions. Error bars represent 95% confi￾dence intervals. 4.1 Search-enabled conditions shifted accuracy between modalities Enabling search was associated with different modality and accuracy relationships across both benchmarks. Without search, chat UI responses were less accurate than API responses. On both benchmarks with search disabled, chat UI responses w… view at source ↗
Figure 2
Figure 2. Within-modality and between-modality re￾sponse disagreement across benchmarks and search con￾ditions. Between-modality disagreement consistently ex￾ceeded within-modality run-to-run disagreement. Error bars represent 95% CIs. For BBQ, the gap between within- and between-modality disagreement was consistent across search conditions. With￾out search, within-modality disagreement for API responses was 8.9% (CI [5.7, 12… view at source ↗
Figure 3
Figure 3. Example benchmark prompt and responses. In this example from our data, the same SafetyBench prompt (A) produced different responses in both the sources cited and the explanation offered, shown in panels (B) and (C), even though both answers would be judged equally by typical benchmarking standards. Jaccard similarity Cosine similarity BBQ SafetyBench no−search search no−search search no−search search no−search searc… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Lexical (Jaccard) and semantic (cosine) similarity by benchmark, search condition, and pair type. Left panel: mean pairwise Jaccard similarity. Right panel: mean cosine similarity of sentence embeddings. Rows: BBQ (top), SafetyBench (bottom). Between-modality similarit…
Figure 5
Figure 5. Figure 5: Source overlap between API and chat UI cita￾tions. Bars show the share of sources cited only by the chat UI (purple), by both modalities (green), or only by the API (blue), at the URL- and doman- levels. 4.5 Abstention was inconsistent across modalities and independent…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

65 extracted references · 22 canonical work pages

  1. [1]

    2026 , note=

    Byeol, Ahn , title=. 2026 , note=

    2026

  2. [2]

    Findings of the

    Parrish, Alicia and Chen, Angelica and Nangia, Nikita and Padmakumar, Vishakh and Phang, Jason and Thompson, Jana and Htut, Phu Mon and Bowman, Samuel , editor =. Findings of the

    work page

  3. [3]

    Proceedings of the 62nd

    Zhang, Zhexin and Lei, Leqi and Wu, Lindong and Sun, Rui and Huang, Yongkang and Long, Chong and Liu, Xiao and Lei, Xuanyu and Tang, Jie and Huang, Minlie , editor =. Proceedings of the 62nd

    work page

  4. [4]

    Proceedings of the 10th SIGHAN Workshop on Chinese Language Processing (SIGHAN-10) , pages=

    Causalbench: A comprehensive benchmark for evaluating causal reasoning capabilities of large language models , author=. Proceedings of the 10th SIGHAN Workshop on Chinese Language Processing (SIGHAN-10) , pages=

    work page

  5. [5]

    arXiv preprint arXiv:2501.14851 , year=

    Justlogic: A comprehensive benchmark for evaluating deductive reasoning in large language models , author=. arXiv preprint arXiv:2501.14851 , year=

    arXiv

  6. [6]

    Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Nphardeval: Dynamic benchmark on reasoning ability of large language models via complexity classes , author=. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    work page

  7. [7]

    arXiv preprint arXiv:2505.02735 , year=

    Formalmath: Benchmarking formal mathematical reasoning of large language models , author=. arXiv preprint arXiv:2505.02735 , year=

    arXiv

  8. [8]

    European Semantic Web Conference , pages=

    Evaluating language models for knowledge base completion , author=. European Semantic Web Conference , pages=. 2023 , organization=

    work page 2023

Show all 65 references
  1. [9]

    Transactions of the Association for Computational Linguistics , volume=

    Natural questions: a benchmark for question answering research , author=. Transactions of the Association for Computational Linguistics , volume=. 2019 , publisher=

    2019

  2. [10]

    Proceedings of the 60th annual meeting of the association for computational linguistics (volume 1: long papers) , pages=

    Truthfulqa: Measuring how models mimic human falsehoods , author=. Proceedings of the 60th annual meeting of the association for computational linguistics (volume 1: long papers) , pages=

  3. [11]

    arXiv preprint arXiv:2107.03374 , year=

    Evaluating large language models trained on code , author=. arXiv preprint arXiv:2107.03374 , year=

    arXiv

  4. [12]

    Proceedings of the 2021 conference of the North American chapter of the association for computational linguistics: human language technologies , pages=

    Are NLP models really able to solve simple math word problems? , author=. Proceedings of the 2021 conference of the North American chapter of the association for computational linguistics: human language technologies , pages=

    2021

  5. [13]

    Findings of the association for computational linguistics: EMNLP 2020 , pages=

    Realtoxicityprompts: Evaluating neural toxic degeneration in language models , author=. Findings of the association for computational linguistics: EMNLP 2020 , pages=

    2020

  6. [14]

    Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing , pages=

    Why should adversarial perturbations be imperceptible? rethink the research paradigm in adversarial NLP , author=. Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing , pages=

    2022

  7. [15]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    T2isafety: Benchmark for assessing fairness, toxicity, and privacy in image generation , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  8. [16]

    C row S -Pairs: A Challenge Dataset for Measuring Social Biases in Masked Language Models

    Nangia, Nikita and Vania, Clara and Bhalerao, Rasika and Bowman, Samuel R. C row S -Pairs: A Challenge Dataset for Measuring Social Biases in Masked Language Models. Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP). 2020. doi:10.18...

    2020 doi

  9. [17]

    Transactions on Machine Learning Research , doi =

    Holistic evaluation of language models , author=. Transactions on Machine Learning Research , doi =

  10. [18]

    ACM transactions on intelligent systems and technology , volume=

    A survey on evaluation of large language models , author=. ACM transactions on intelligent systems and technology , volume=. 2024 , publisher=

    2024

  11. [19]

    Transactions on machine learning research , year=

    Beyond the imitation game: Quantifying and extrapolating the capabilities of language models , author=. Transactions on machine learning research , year=

  12. [20]

    arXiv preprint arXiv:2009.03300 , year=

    Measuring massive multitask language understanding , author=. arXiv preprint arXiv:2009.03300 , year=

    2009 arXiv

  13. [21]

    Findings of the Association for Computational Linguistics: NAACL 2024 , pages=

    Large language models sensitivity to the order of options in multiple-choice questions , author=. Findings of the Association for Computational Linguistics: NAACL 2024 , pages=

    2024

  14. [22]

    P ro SA : Assessing and Understanding the Prompt Sensitivity of LLM s

    Zhuo, Jingming and Zhang, Songyang and Fang, Xinyu and Duan, Haodong and Lin, Dahua and Chen, Kai. P ro SA : Assessing and Understanding the Prompt Sensitivity of LLM s. Findings of the Association for Computational Linguistics: EMNLP 2024. 2024. doi:10.18653/v1/2024.findings-...

    2024 doi

  15. [23]

    Advances in Neural Information Processing Systems , volume=

    Make your llm fully utilize the context , author=. Advances in Neural Information Processing Systems , volume=

  16. [24]

    Proceedings of the 1st ACM workshop on large AI systems and models with privacy and safety analysis , pages=

    Promptrobust: Towards evaluating the robustness of large language models on adversarial prompts , author=. Proceedings of the 1st ACM workshop on large AI systems and models with privacy and safety analysis , pages=

  17. [25]

    arXiv preprint arXiv:2401.08329 , year=

    Understanding user experience in large language model interactions , author=. arXiv preprint arXiv:2401.08329 , year=

    arXiv

  18. [26]

    arXiv preprint arXiv:2403.04132 , year=

    Chatbot arena: An open platform for evaluating llms by human preference , author=. arXiv preprint arXiv:2403.04132 , year=

    arXiv

  19. [27]

    Advances in Neural Information Processing Systems , volume=

    Beyond prompts: Dynamic conversational benchmarking of large language models , author=. Advances in Neural Information Processing Systems , volume=

  20. [28]

    Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Chatbench: From static benchmarks to human-ai evaluation , author=. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  21. [29]

    arXiv preprint arXiv:2509.19364 , year=

    The inadequacy of offline llm evaluations: A need to account for personalization in model behavior , author=. arXiv preprint arXiv:2509.19364 , year=

  22. [30]

    A multitask, multilingual, multimodal evaluation of chatgpt on reasoning, hallucination, and interactivity , author=. Proceedings of the 13th international joint conference on natural language processing and the 3rd conference of the asia-pacific chapter of the association for...

  23. [31]

    2023 , isbn =

    Li, Xinzhe and Liu, Ming and Gao, Shang and Buntine, Wray , title =. 2023 , isbn =. doi:10.24963/ijcai.2023/749 , booktitle =

    2023 doi

  24. [32]

    arXiv preprint arXiv:2302.12095 , year=

    On the robustness of chatgpt: An adversarial and out-of-distribution perspective , author=. arXiv preprint arXiv:2302.12095 , year=

    arXiv

  25. [33]

    Ask Again, Then Fail: Large Language Models' Vacillations in Judgment

    Xie, Qiming and Wang, Zengzhi and Feng, Yi and Xia, Rui. Ask Again, Then Fail: Large Language Models' Vacillations in Judgment. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024. doi:10.18653/v1/2024.acl-long.577

    2024 doi

  26. [34]

    S iren ' s Song in the AI Ocean: A Survey on Hallucination in Large Language Models

    Zhang, Yue and Li, Yafu and Cui, Leyang and Cai, Deng and Liu, Lemao and Fu, Tingchen and Huang, Xinting and Zhao, Enbo and Zhang, Yu and Chen, Yulong and Wang, Longyue and Luu, Anh Tuan and Bi, Wei and Shi, Freda and Shi, Shuming. S iren ' s Song in the AI Ocean: A Survey on ...

    2025 doi

  27. [35]

    Proceedings of the 2021 ACM conference on fairness, accountability, and transparency , pages=

    Bold: Dataset and metrics for measuring biases in open-ended language generation , author=. Proceedings of the 2021 ACM conference on fairness, accountability, and transparency , pages=

    2021

  28. [36]

    arXiv preprint arXiv:2304.03738 , year=

    Should chatgpt be biased? challenges and risks of bias in large language models , author=. arXiv preprint arXiv:2304.03738 , year=

    arXiv

  29. [37]

    arXiv preprint arXiv:2604.06188 , year=

    LLM Spirals of Delusion: A Benchmarking Audit Study of AI Chatbot Interfaces , author=. arXiv preprint arXiv:2604.06188 , year=

    arXiv

  30. [38]

    Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society , number=

    Can we trust ai benchmarks? an interdisciplinary review of current issues in ai evaluation , author=. Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society , number=

  31. [39]

    Practices for

    CAISI , year = 2026, number =. Practices for

    2026

  32. [40]

    CAISI , institution =

  33. [41]

    arXiv preprint arXiv:2008.02275 , year=

    Aligning ai with shared human values , author=. arXiv preprint arXiv:2008.02275 , year=

    2008 arXiv

  34. [42]

    arXiv preprint arXiv:2503.05336 , year=

    Toward an evaluation science for generative ai systems , author=. arXiv preprint arXiv:2503.05336 , year=

    arXiv

  35. [43]

    Science , volume=

    Rethink reporting of evaluation results in AI , author=. Science , volume=. 2023 , publisher=

    2023

  36. [44]

    Burnell, Ryan and Burden, John and Rutar, Danaja and Voudouris, Konstantinos and Cheke, Lucy and. Not a. Proceedings of the. doi:10.24963/ijcai.2022/392 , isbn =

    2022 doi

  37. [45]

    arXiv preprint arXiv:2412.03597 , year=

    The Vulnerability of Language Model Benchmarks: Do They Accurately Reflect True LLM Performance? , author=. arXiv preprint arXiv:2412.03597 , year=

    arXiv

  38. [46]

    Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    When benchmarks are targets: Revealing the sensitivity of large language model leaderboards , author=. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  39. [47]

    and Paullada, Amandalynne and Denton, Emily and Hanna, Alex , year = 2021, month = nov, eprint =

    Raji, Inioluwa Deborah and Bender, Emily M. and Paullada, Amandalynne and Denton, Emily and Hanna, Alex , year = 2021, month = nov, eprint =. doi:10.48550/arXiv.2111.15366 , archiveprefix =

    2021 doi arXiv

  40. [48]

    Proceedings of the 2025

    Nalbandyan, Grigor and Shahbazyan, Rima and Bakhturina, Evelina , editor =. Proceedings of the 2025. doi:10.18653/v1/2025.naacl-industry.39 , isbn =

    2025 doi

  41. [49]

    Rethinking

    Ganesh, Prakhar and Shokri, Reza and Farnadi, Golnoosh , editor =. Rethinking. Proceedings of the 19th. doi:10.18653/v1/2026.eacl-long.327 , isbn =

    2026 doi

  42. [50]

    arXiv preprint arXiv:2502.00561 , year=

    Position: Evaluating generative ai systems is a social science measurement challenge , author=. arXiv preprint arXiv:2502.00561 , year=

    arXiv

  43. [51]

    Outsider

    Raji, Inioluwa Deborah and Xu, Peggy and Honigsberg, Colleen and Ho, Daniel , year = 2022, month = jul, pages =. Outsider. Proceedings of the 2022. doi:10.1145/3514094.3534181 , isbn =

    2022

  44. [52]

    arXiv preprint arXiv:2310.11986 , year=

    Sociotechnical safety evaluation of generative ai systems , author=. arXiv preprint arXiv:2310.11986 , year=

    arXiv

  45. [53]

    arXiv preprint arXiv:2405.14782 , year=

    Lessons from the trenches on reproducible evaluation of language models , author=. arXiv preprint arXiv:2405.14782 , year=

    arXiv

  46. [54]

    Advances in Neural Information Processing Systems , volume=

    Betterbench: Assessing ai benchmarks, uncovering issues, and establishing best practices , author=. Advances in Neural Information Processing Systems , volume=

  47. [55]

    Proceedings of the 29th international conference on computational linguistics , pages=

    BECEL: Benchmark for consistency evaluation of language models , author=. Proceedings of the 29th international conference on computational linguistics , pages=

  48. [56]

    International Conference on Learning Representations , volume=

    Swe-bench: Can language models resolve real-world github issues? , author=. International Conference on Learning Representations , volume=

  49. [57]

    Findings of the Association for Computational Linguistics: ACL 2024 , pages=

    Freshllms: Refreshing large language models with search engine augmentation , author=. Findings of the Association for Computational Linguistics: ACL 2024 , pages=

    2024

  50. [58]

    arXiv preprint arXiv:2504.12516 , year=

    Browsecomp: A simple yet challenging benchmark for browsing agents , author=. arXiv preprint arXiv:2504.12516 , year=

    arXiv

  51. [59]

    and Darrell, Trevor and Norouzi, Narges and Gonzalez, Joseph E

    Miroyan, Mihran and Wu, Tsung-Han and King, Logan and Li, Tianle and Pan, Jiayi and Hu, Xinyan and Chiang, Wei-Lin and Angelopoulos, Anastasios N. and Darrell, Trevor and Norouzi, Narges and Gonzalez, Joseph E. , year = 2026, month = mar, number =. Search. doi:10.48550/arXiv.2...

    2026 doi

  52. [60]

    Kale, Sahil , year = 2025, month = nov, number =. Look. doi:10.48550/arXiv.2511.18931 , archiveprefix =. 2511.18931 , primaryclass =

    2025 doi

  53. [61]

    Sacred or

    Atif, Farah and Askarbekuly, Nursultan and Darwish, Kareem and Choudhury, Monojit , year = 2025, month = oct, journal =. Sacred or

    2025

  54. [62]

    , year = 2025, month = jun, number =

    Kirichenko, Polina and Ibrahim, Mark and Chaudhuri, Kamalika and Bell, Samuel J. , year = 2025, month = jun, number =. doi:10.48550/arXiv.2506.09038 , archiveprefix =. 2506.09038 , primaryclass =

    2025 doi arXiv

  55. [63]

    and Metaxa, Dana

    Proebsting, Grace and Anigboro, Oghenefejiro Isaacs and Crawford, Charlie M. and Metaxa, Dana. Identity-Related. Proceedings of the 5th. doi:10.1145/3757887.3763010 , isbn =

  56. [64]

    Proceedings of the 2020 conference on fairness, accountability, and transparency , pages=

    Closing the AI accountability gap: Defining an end-to-end framework for internal algorithmic auditing , author=. Proceedings of the 2020 conference on fairness, accountability, and transparency , pages=

    2020

  57. [65]

    arXiv preprint arXiv:2503.16861 , year=

    In-house evaluation is not enough: Towards robust third-party flaw disclosure for general-purpose ai , author=. arXiv preprint arXiv:2503.16861 , year=

    arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.