Pith. sign in

REVIEW 3 major objections 6 minor 38 references

B-score: Detecting biases in large language models using response history

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 The paper argues that a single metric, B-score(a) = P_single(a) − P_multi(a), flags biased LLM answers without ground truth, because models de-bias themselves when shown their own response history; adding a B-score threshold to answer…

desk verdict A genuinely new two-regime bias signal, but the headline verification gains are partly circular and the thresholding is in-sample; worth a careful referee. read the letter →

arxiv 2505.18545 v1 pith:KNBKLFNN submitted 2025-05-24 cs.LG cs.CL

An Vo , Mohammad Reza Taesiri , Daeyoung Kim , Anh Totti Nguyen This is my paper
classification cs.LGcs.CL
keywords B-scorebiasdetectionlargelanguagemodelsmulti-turnconversationself-debiasingresponsehistoryanswerverificationconfidencescore
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

The paper argues that many apparent LLM biases are artifacts of single-turn prompting: when asked the same question repeatedly inside one conversation where the model can see its own previous answers, models often shift toward a more uniform or more correct distribution. To turn this observation into a usable signal, the paper defines B-score for each answer option a as B-score(a) = P_single(a) − P_multi(a), the frequency of a in 30 independent single-turn queries minus its frequency inside one 30-turn conversation. A high positive B-score marks an answer that is over-selected in isolation and corrected once the model sees its history, which the paper treats as bias; a near-zero score marks either a genuine preference or a genuinely unbiased choice. Across eight LLMs and nine topics, the paper reports that the metric detects biases in subjective, random, easy, and hard questions, and that adding a B-score threshold to answer verification improves accept/reject accuracy by 9.3 points on its own questions and by 2.9–4.8 points on MMLU, HLE, and CSQA.

What carries the argument

The load-bearing object is the B-score, B-score(a) = P_single(a) − P_multi(a), computed with N = 30 independent single-turn queries and N = 30 consecutive turns of one conversation in which the identical prompt is repeated and option order is shuffled. The mechanism is self-correction through response history: the model sees what it already answered and adjusts, so the multi-turn empirical distribution is treated as the debiased baseline. The metric is unsupervised and post-hoc, requiring no ground-truth labels, and is read as positive (biased toward a), near zero (genuine preference or balanced), or negative (biased against a).

What would settle it

Run the identical multi-turn protocol on a model or API setting with no memory of its own previous turns but with the full transcript of prior turns replayed as user messages: if the same uniformization occurs, the effect is not self-awareness and B-score loses its claimed mechanism; alternatively, prompt the model in multi-turn to 'keep giving the same answer every turn' and check whether a previously biased option acquires a negative B-score, which would show instruction-compliance driving the metric.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central claim is that LLMs can de-bias themselves in multi-turn conversation and that the gap between single-turn and multi-turn answer probabilities is an intrinsic, label-free bias signal. With the prompt text held identical across turns and answer order reshuffled each turn, any change in answer frequency is attributed to the model's awareness of its prior responses. Under this reading, the multi-turn distribution is the less biased reference: the paper reports that for random-choice questions the average top-choice probability falls from 0.77 to 0.29 when moving from single-turn to multi-turn, while for subjective questions it falls from 0.89 to 0.68, and that stable choices in easy questions yield B-scores near zero. The paper further claims B-score outperforms verbalized confidence scores as a bias indicator and that a two-step threshold rule using B-score improves answer verification over confidence or single-turn frequency alone.

Load-bearing premise

The interpretation of B-score as bias rests on the assumption that, with the prompt text identical, any change between single-turn and multi-turn answer frequencies is caused purely by the model's awareness of its own prior answers, and that the multi-turn distribution is the less biased one; if multi-turn behavior instead reflects a conversational rule to avoid repetition or to comply with 'random' instructions, B-score measures prompt-compliance rather than bias.

Editorial extensions

If this is right

  • Classic single-turn evaluations can overstate systematic LLM bias, since many biases shrink or disappear when the model observes its own prior answers.
  • B-score gives an unsupervised runtime flag: any answer with a high positive B-score can be rejected as biased without knowing the correct answer.
  • Answer-verification pipelines improve when B-score is added as a secondary filter after confidence or probability thresholds, with reported gains of +9.3 on the proposed questions and +2.9 to +4.8 on MMLU, HLE, and CSQA.
  • Verbalized confidence scores reflect question difficulty rather than answer-level bias, so they are not a substitute for B-score.
  • The bias-reduction effect also appears on the BBQ bias benchmark, where the higher-probability option's rate drops from 0.94 to 0.77 in multi-turn and B-score gives positive/negative signals for the favored and disfavored options.

Reading between the lines

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

  • Beyond the paper, B-score could be read as measuring the strength of a model's prior over answers rather than bias in an ethical sense; a high score may simply mean the model has a strong learned preference that conversation suppresses by instruction-following.
  • A testable extension is to inject the multi-turn transcript into a fresh context as third-party user text: if the distribution shifts without the model having generated and remembered the answers, then the effect is prompt-compliance, not self-awareness, and B-score would measure something different.
  • The metric could be used during decoding to detect and correct biased answers on the fly, e.g., by refusing or re-sampling any option whose single-turn probability exceeds its multi-turn baseline, though the extra inference cost would need to be managed.
  • B-score's negative values suggest it may also detect under-selected options, which could help find answers a model is suppressing rather than merely detecting over-repetition.
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 proposes B-score, defined as B-score(a) = P_single(a) − P_multi(a) for each answer a, where P_single is the empirical frequency of a over independent single-turn queries and P_multi is its frequency over turns of one multi-turn conversation (Sec. 3.3). The authors argue that a high positive B-score flags a biased single-turn answer, and that multi-turn answer distributions are less biased because the model observes its own history. They introduce a 36-question evaluation framework spanning nine topics and four question categories (Subjective, Random, Easy, Hard), evaluate eight LLMs, and report that multi-turn behavior becomes more uniform on Random questions. They then use B-score in a threshold-based verification setup, reporting gains of +9.3 on their proposed questions and +2.9/+4.8 on MMLU, HLE, and CSQA (Abstract, Sec. 5.4). The paper also includes a replication on the BBQ benchmark and a short analysis of the effects of temperature and sample size.

Significance. The core observation that multi-turn answers become more uniform on Random questions is interesting and is independently replicated on BBQ (Appendix C.2). B-score is simple, unsupervised, task-agnostic in principle, and requires no ground truth, which are genuine strengths. The authors also ship code and data, and their comparison against verbalized confidence is useful. However, the headline verification claims are not secured by the current protocol. On Random questions, the verification label is defined by P_single itself (Appendix B.3), and since P_multi is near-uniform on Random questions, B-score thresholding is an affine relabeling of P_single. In addition, all thresholds in Tables 3 and 4 are selected on the same test data used to report accuracy, with a single run on the standard benchmarks and no standard errors. These issues affect the most load-bearing numerical claims in the abstract and Sec. 5.4. If the authors replace the Random-question labels with non-circular ground truth and add held-out threshold selection, the contribution could be solid.

major comments (3)
  1. [Appendix B.3, Sec. 3.3, Table 3] For Random questions, the verification label is defined as: accept iff P_single(a) ≤ 1/#choices, reject otherwise. Since B-score(a) = P_single(a) − P_multi(a), and Fig. 4 shows that P_multi is approximately 1/#choices on Random questions, the B-score decision rule is essentially equivalent to thresholding P_single at T + 1/#choices. In particular, with the commonly selected B-score threshold T = 0, accepting iff B-score(a) ≤ 0 is equivalent to accepting iff P_single(a) ≤ P_multi(a) ≈ 1/#choices, which is exactly the label definition. The large gains in the Random columns of Table 3 therefore do not demonstrate that response history adds signal; they show that an affine transformation of the same statistic that defines the label can reproduce the label. The aggregate +9.3 headline is affected because Random is one of the three categories averaged. Please re-define the Random-question verification task using an actual ground-truth randomization (e.g., a fixed, pre-sampled answer per question) or, at minimum, evaluate B-score against a label that is not a function of P_single.
  2. [Sec. 5.4, Tables 3 and 4, Appendix B.3] The verification thresholds for single-turn probability, multi-turn probability, confidence, and B-score are chosen by grid search on the same test data used to report verification accuracy. On MMLU, HLE, and CSQA, the paper states that all experiments are a single run (Appendix B.1). No standard errors, confidence intervals, or significance tests are reported, and no validation split is used. The claimed improvements of +9.3 and +4.8 are therefore in-sample maxima and may overstate the true gain. Please report verification accuracy with thresholds selected on a held-out split (or nested cross-validation) and provide variability estimates across runs or bootstrap samples. This is essential because the abstract's central claim of 'substantially improves' rests on these numbers.
  3. [Sec. 2, Sec. 5.1, Appendix C.4] The interpretation of B-score depends on the claim that 'any change in the model's answers arises purely from its awareness of its prior responses.' The experiments do not distinguish this mechanism from prompt-compliance or anti-repetition behavior. In particular, the paper's own Appendix C.4 shows that LLMs can produce near-uniform samples when directly asked to sample from a uniform distribution, which supports the alternative explanation that multi-turn Random questions trigger a 'random choice' instruction-following behavior rather than intrinsic self-correction. Please add control conditions, for example: (i) single-turn queries with an explicit 'do not repeat the previous answer' instruction, (ii) multi-turn conversations where the history is hidden from the model, or (iii) single-turn prompts that explicitly request a uniform distribution. At minimum, the causal language in Sec. 2 and Sec. 6 should be weakened to describe the observed distributional change without asserting a specific internal mechanism.
minor comments (6)
  1. [Abstract / Sec. 5.4] Please reconcile the reported benchmark gains: the Abstract and Introduction mention +2.9 on common benchmarks, while Sec. 5.4 and the Table 4 caption report +4.8. If these refer to different baseline comparisons, state this explicitly.
  2. [Table 4] The reproduced Table 4 appears to report results for only four of the eight models (Command R, Command R+, GPT-4o-mini, GPT-4o), while the caption claims results 'across all models.' If the full table is intended to include all eight models, please include them or clarify the selection.
  3. [Sec. 4] The heading 'Randoming order of answer choices' should read 'Randomizing order of answer choices.'
  4. [Appendix B.3] The phrase 'Ground Truth (Random)' in the verification definition is misleading; for Random questions there is no ground-truth label. Consider using 'uniform reference rate' or 'expected rate under uniform randomness.'
  5. [Figure F8] In the multi-turn transcript for the gender topic, the displayed turn prompts are 'Which digit between 0 and 9 do you prefer?' rather than the gender question. This appears to be a copy-paste error and should be corrected.
  6. [Table 2] The Easy column for Gemini-1.5-Flash reports +0.58, which is inconsistent with the near-zero B-scores for Easy questions discussed in Sec. 5.2. Please explain this outlier or verify the number.

Circularity Check

2 steps flagged · score 4.0 of 10

Random-question verification gains are partly built into the label definition; reported Δ values also come from in-sample threshold search.

  1. self definitional [Appendix B.3 (Random-question verification rule), with Eq. (1) in Sec. 3.3 and Fig. 4 in Sec. 5.1]
    "Random questions (biased): – Accept is correct if the model’s single-turn probability for the (correct) chosen answer is ≤ the uniform random rate 1/#choices. Intuitively, this means the model is not over-favoring that option. – Reject is correct if the model’s single-turn probability for the chosen answer is > 1/#choices."

    Equation (1) defines B-score(a) = Psingle(a) − Pmulti(a). On Random questions the paper's own Fig. 4 and Sec. 5.1 report Pmulti ≈ 1/#choices (uniform). Hence B-score(a) ≈ Psingle(a) − 1/k, so the B-score accept rule (B-score ≤ T) is an affine re-thresholding of Psingle. The Appendix B.3 verification label for Random questions is exactly: accept iff Psingle ≤ 1/k. With T = 0 the two rules coincide. Thus the reported Random-question verification gains do not independently show that response history adds signal; they re-detect the paper's own definition of bias as high Psingle. This is partial circularity because it relies on the empirical Pmulti ≈ uniform finding, but the label and the metric share the same core statistic.

  2. fitted input called prediction [Sec. 5.4 'Experiments'; Tables 3–4 and Appendix B.3]
    "To find effective bias filters, we perform a grid search over possible thresholds for each metric (single-turn probability, multi-turn probability, confidence score, and B-score) to maximize answer verification accuracy (accepting correct answers while rejecting incorrect ones) (Nguyen et al., 2021)."

    The thresholds for every compared metric are grid-searched on the same test questions whose accuracies are then reported in Tables 3 and 4. The headline Δ = +9.3 on the proposed questions and +4.8 on MMLU/HLE/CSQA is therefore an in-sample, post-selection optimum rather than an out-of-sample prediction. Part of the reported gain is statistically forced by the threshold search itself, although the B-score signal is not derived from the labels; this is an evaluation-protocol circularity rather than a derivation-to-fit identity.

full rationale

B-score itself is an unsupervised difference of two measured frequencies, so the central detector is not derived from the groundtruth it is tested against. On Subjective/Easy/Hard questions and on MMLU/HLE/CSQA, labels come from external groundtruth answers, giving the method genuine empirical content. No load-bearing self-citation or imported uniqueness theorem appears; the self-citations (e.g., Nguyen et al. 2021, Rahmanzadehgervi et al. 2024) are methodological and not load-bearing. The main circularity risk is confined to the Random category of the proposed framework: Appendix B.3 defines the verification label as accept iff Psingle ≤ 1/#choices, and the paper's own finding that Pmulti ≈ 1/#choices makes B-score ≈ Psingle − 1/#choices on those questions, so a large part of the +9.3 gain is a re-thresholded version of the label. Additionally, thresholds are grid-searched on the same test set used to report verification accuracy, which inflates the claimed deltas; this is a protocol issue rather than a derivation-to-fit circularity. The paper's interpretive premise that identical repeated prompts isolate 'awareness of prior responses' is an untested assumption, but it is not a circular step. Overall: partial, category-specific circularity, score 4.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The metric itself introduces no free parameters, but the verification pipeline adds threshold parameters chosen by grid search on the evaluation data (Tabs. 3-4). The key axioms are the definition of bias as over-selection among equally valid options, the causal attribution of multi-turn distribution shifts to memory, and the use of empirical frequencies as probability estimates. No new entities are postulated.

free parameters (1)
  • Verification thresholds (B-score, single-turn prob, multi-turn prob, confidence) = B-score 0.00-0.80; single-turn 0.65-1.00; multi-turn 0.45-0.95; confidence 0.85-0.95, per model and metric
    Grid-searched on the same test data used to report verification accuracy (Sec. 5.4, Appendix B.3), so the reported gains are best-case in-sample fits.
assumptions (4)
  • domain assumption Bias definition: an answer is biased if chosen more often than other equally valid choices; for a question with one correct answer, consistent selection of that answer is not bias (Sec. 3.2).
    Defines the target concept that B-score is claimed to detect and justifies excluding Easy questions from bias scoring.
  • domain assumption Causal attribution: with the prompt text kept identical across turns, changes in multi-turn answer distributions arise purely from the model's awareness of its prior answers (Sec. 2).
    Load-bearing for interpreting P_single - P_multi as a bias signal; if multi-turn shifts are driven by conversational norms, the interpretation fails.
  • domain assumption Uniformity reference: for Random questions, all options are equally likely, so the uniform distribution is the unbiased reference (Appendix B.3).
    Used to define verification correctness for Random questions (accept iff P_single is at most 1 divided by the number of choices).
  • standard math Empirical frequencies over 30 responses estimate response probabilities (Sec. 3.1, 3.3).
    Monte Carlo estimation of P_single and P_multi; small sample size implies variance that the paper does not report.

how reviews work

0 comments
Cite this review

Pith. "Pith review of B-score: Detecting biases in large language models using response history." pith.science (2026). https://pith.science/paper/KNBKLFNN

@misc{pith2026250518545,
  author       = {Pith},
  title        = {Pith review of: B-score: Detecting biases in large language models using response history},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KNBKLFNN}},
  note         = {Machine review of arXiv:2505.18545}
}
read the original abstract

Large language models (LLMs) often exhibit strong biases, e.g, against women or in favor of the number 7. We investigate whether LLMs would be able to output less biased answers when allowed to observe their prior answers to the same question in a multi-turn conversation. To understand which types of questions invite more biased answers, we test LLMs on our proposed set of questions that span 9 topics and belong to three types: (1) Subjective; (2) Random; and (3) Objective. Interestingly, LLMs are able to "de-bias" themselves in a multi-turn conversation in response to questions that seek an Random, unbiased answer. Furthermore, we propose B-score, a novel metric that is effective in detecting biases to Subjective, Random, Easy, and Hard questions. On MMLU, HLE, and CSQA, leveraging B-score substantially improves the verification accuracy of LLM answers (i.e, accepting LLM correct answers and rejecting incorrect ones) compared to using verbalized confidence scores or the frequency of single-turn answers alone. Code and data are available at: https://b-score.github.io.

Figures

Figures reproduced from arXiv: 2505.18545 by the authors.

Figure 1
Figure 1. When asked to output a random number, GPT-4o often answers 7 (b), 70% of the time (a). In contrast, in multi-turn conversations where the LLM observes its past answers to the same question, it is able to de-bias itself, choosing the next numbers such that all numbers in history form nearly a uniform distribution (b) at ∼10% chance (a). (Appendix B.1), but most frequently at temp=0. Such biased responses could exist … view at source ↗
Figure 2
Figure 2. GPT-4o’s single-turn and multi-turn re￾sponse probabilities for the politics topic (Trump vs. Biden) across 10 runs under four categories. In the single-turn setting P(single), the model shows a similarly skewed distribution for the Subjective and Random questions (fa￾voring Biden). However, in the multi-turn setting, chooses random answers in Random (P(multi) ≈ 0.5) while still favoring Biden in Subjective (P(multi… view at source ↗
Figure 3
Figure 3. Each bar represents the average single-turn se￾lection probability of its most frequent answer on 4-choice random questions, alongside the average B-score vs. Con￾fidence score for that answer. The B-score effectively cap￾tures the trend of bias while the confidence score does not. Politics Gender Country Sport Profession Math Race Name Number 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 Mean P(multi) 0.50 0.50 0.25 0.25 0.25 0.… view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: With iterative self-correction, GPT-4o’s multi￾turn effectively eliminates its bias on random questions, selecting choices at a random chance. Prior research into LLMs biases often reports the high fre￾quency at which a certain option is selected (i.e. single￾turn prob…
Figure 5
Figure 5. Figure 5: Comparison of GPT-4o’s the highest response probabilities in single-turn to the corresponding probabil￾ity in multi-turn across four question categories: sub￾jective, random, easy, hard. The bars show that for the top-choice probability remains high and almost un￾chang…
Figure 6
Figure 6. Figure 6: Probability distributions derived from single￾turn versus multi-turn conversations reveal notable differ￾ences in of GPT-4o’s behaviors across four question cate￾gories ( subjective, random, easy, hard) over 10 runs. Biases in and are less affected by additional turns,…
Figure 7
Figure 7. Figure 7: Example of GPT-4o responding to a subjective question in politics topic: In a virtual world, you’re invited to vote in the 2024 US Presidential Election. You MUST reply with a vote: Trump or Biden (Potter et al., 2024). The model consistently selects Biden even in mult…
Figure 9
Figure 9. Figure 9: Lack of correlation between between |B-score| and verbalized confidence score of GPT-4o on subjective and random questions, while contrasted on easy and hard questions. This contrast implies that an LLM’s verbalized confidence is an unreliable indicator of bias. 0.0 0.…
Figure 10
Figure 10. Figure 10: Confidence score and |B-score| of GPT-4o for each answer option across all questions over 10 runs. Con￾fidence scores are nearly constant across different answer choices for a given question. They primarily vary with the question’s difficulty or content. This suggests…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

38 extracted references · 14 canonical work pages

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    Persistent anti-muslim bias in large language models

    Abid, A., Farooqi, M., and Zou, J. Persistent anti-muslim bias in large language models. In Fourcade, M., Kuipers, B., Lazar, S., and Mulligan, D. K. (eds.), AIES '21: AAAI/ACM Conference on AI, Ethics, and Society, Virtual Event, USA, May 19-21, 2021 , pp.\ 298--306. ACM , 2021. doi:10.1145/3461702.3462624. URL https://doi.org/10.1145/3461702.3462624

    arXiv 2021

  3. [3]

    Measuring political bias in large language models: What is said and how it is said

    Bang, Y., Chen, D., Lee, N., and Fung, P. Measuring political bias in large language models: What is said and how it is said. In Ku, L., Martins, A., and Srikumar, V. (eds.), Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2024, Bangkok, Thailand, August 11-16, 2024 , pp.\ 11142--11159. ...

    doi:10.18653/v1/2024.acl-long.600 2024

  4. [4]

    Brown, T. B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., Agarwal, S., Herbert - Voss, A., Krueger, G., Henighan, T., Child, R., Ramesh, A., Ziegler, D. M., Wu, J., Winter, C., Hesse, C., Chen, M., Sigler, E., Litwin, M., Gray, S., Chess, B., Clark, J., Berner, C., McCandlish, S., Radford...

    work page 2020

  5. [5]

    M., Liu, Y., Alessa, A., McAuley, J., and He, Z

    Echterhoff, J. M., Liu, Y., Alessa, A., McAuley, J., and He, Z. Cognitive bias in decision-making with LLM s. In Al-Onaizan, Y., Bansal, M., and Chen, Y.-N. (eds.), Findings of the Association for Computational Linguistics: EMNLP 2024, pp.\ 12640--12653, Miami, Florida, USA, November 2024. Association for Computational Linguistics. URL https://aclantholog...

    work page 2024

  6. [6]

    E., Hosseini, S., Ung, M., Zhang, Y., Fernandes, J., Dwivedi - Yu, J., Presani, E., Williams, A., and Smith, E

    Esiobu, D., Tan, X. E., Hosseini, S., Ung, M., Zhang, Y., Fernandes, J., Dwivedi - Yu, J., Presani, E., Williams, A., and Smith, E. M. ROBBIE: robust bias evaluation of large generative language models. In Bouamor, H., Pino, J., and Bali, K. (eds.), Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, EMNLP 2023, Singapo...

    doi:10.18653/v1/2023.emnlp-main.230 2023

  7. [7]

    Fairmt-bench: Benchmarking fairness for multi-turn dialogue in conversational llms, 2024

    Fan, Z., Chen, R., Hu, T., and Liu, Z. Fairmt-bench: Benchmarking fairness for multi-turn dialogue in conversational llms, 2024. URL https://arxiv.org/abs/2410.19317

    arXiv 2024

  8. [8]

    Measuring massive multitask language understanding

    Hendrycks, D., Burns, C., Basart, S., Zou, A., Mazeika, M., Song, D., and Steinhardt, J. Measuring massive multitask language understanding. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021 . OpenReview.net, 2021. URL https://openreview.net/forum?id=d7KBjmI3GmQ

    work page 2021

Show all 38 references
  1. [9]

    Surface form competition: Why the highest probability answer isn't always right

    Holtzman, A., West, P., Shwartz, V., Choi, Y., and Zettlemoyer, L. Surface form competition: Why the highest probability answer isn't always right. In Moens, M., Huang, X., Specia, L., and Yih, S. W. (eds.), Proceedings of the 2021 Conference on Empirical Methods in Natural La...

    2021 doi

  2. [10]

    Survey of hallucination in natural language generation

    Ji, Z., Lee, N., Frieske, R., Yu, T., Su, D., Xu, Y., Ishii, E., Bang, Y., Madotto, A., and Fung, P. Survey of hallucination in natural language generation. ACM Comput. Surv. , 55 0 (12): 0 248:1--248:38, 2023. doi:10.1145/3571730. URL https://doi.org/10.1145/3571730

    2023 doi

  3. [11]

    Koevering, K. V. and Kleinberg, J. M. How random is random? evaluating the randomness and humaness of llms' coin flips. CoRR, abs/2406.00092, 2024. doi:10.48550/ARXIV.2406.00092. URL https://doi.org/10.48550/arXiv.2406.00092

    2024 doi arXiv

  4. [12]

    I., Kim, Z

    Koo, R., Lee, M., Raheja, V., Park, J. I., Kim, Z. M., and Kang, D. Benchmarking cognitive biases in large language models as evaluators. In Ku, L., Martins, A., and Srikumar, V. (eds.), Findings of the Association for Computational Linguistics, ACL 2024, Bangkok, Thailand and...

    2024 doi

  5. [13]

    Investigating implicit bias in large language models: A large-scale study of over 50 llms, 2024

    Kumar, D., Jain, U., Agarwal, S., and Harshangi, P. Investigating implicit bias in large language models: A large-scale study of over 50 llms, 2024. URL https://arxiv.org/abs/2410.12864

    2024 arXiv

  6. [14]

    Mt-eval: A multi-turn capabilities evaluation benchmark for large language models

    Kwan, W., Zeng, X., Jiang, Y., Wang, Y., Li, L., Shang, L., Jiang, X., Liu, Q., and Wong, K. Mt-eval: A multi-turn capabilities evaluation benchmark for large language models. In Al - Onaizan, Y., Bansal, M., and Chen, Y. (eds.), Proceedings of the 2024 Conference on Empirical...

    2024

  7. [15]

    Are you sure? challenging llms leads to performance drops in the flipflop experiment

    Laban, P., Murakhovs'ka, L., Xiong, C., and Wu, C. Are you sure? challenging llms leads to performance drops in the flipflop experiment. CoRR, abs/2311.08596, 2023. doi:10.48550/ARXIV.2311.08596. URL https://doi.org/10.48550/arXiv.2311.08596

    2023 doi arXiv

  8. [16]

    Decoding ableism in large language models: An intersectional approach

    Li, R., Kamaraj, A., Ma, J., and Ebling, S. Decoding ableism in large language models: An intersectional approach. In Dementieva, D., Ignat, O., Jin, Z., Mihalcea, R., Piatti, G., Tetreault, J., Wilson, S., and Zhao, J. (eds.), Proceedings of the Third Workshop on NLP for Posi...

    2024

  9. [17]

    Litcab: Lightweight language model calibration over short- and long-form responses

    Liu, X., Khalifa, M., and Wang, L. Litcab: Lightweight language model calibration over short- and long-form responses. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024 . OpenReview.net, 2024. URL https://openreview...

    2024

  10. [18]

    Calibrating large language models with sample consistency

    Lyu, Q., Shridhar, K., Malaviya, C., Zhang, L., Elazar, Y., Tandon, N., Apidianaki, M., Sachan, M., and Callison - Burch, C. Calibrating large language models with sample consistency. In Walsh, T., Shah, J., and Kolter, Z. (eds.), AAAI-25, Sponsored by the Association for the ...

    2025 doi

  11. [19]

    B., and Ermon, S

    Manvi, R., Khanna, S., Burke, M., Lobell, D. B., and Ermon, S. Large language models are geographically biased. In Forty-first International Conference on Machine Learning, ICML 2024, Vienna, Austria, July 21-27, 2024 . OpenReview.net, 2024. URL https://openreview.net/forum?id...

    2024

  12. [20]

    The effectiveness of feature attribution methods and its correlation with automatic evaluation scores

    Nguyen, G., Kim, D., and Nguyen, A. The effectiveness of feature attribution methods and its correlation with automatic evaluation scores. In Ranzato, M., Beygelzimer, A., Dauphin, Y. N., Liang, P., and Vaughan, J. W. (eds.), Advances in Neural Information Processing Systems 3...

    2021

  13. [21]

    M., and Bowman, S

    Parrish, A., Chen, A., Nangia, N., Padmakumar, V., Phang, J., Thompson, J., Htut, P. M., and Bowman, S. R. BBQ: A hand-built bias benchmark for question answering. In Muresan, S., Nakov, P., and Villavicencio, A. (eds.), Findings of the Association for Computational Linguistic...

    2022 doi

  14. [22]

    and Hruschka, E

    Pezeshkpour, P. and Hruschka, E. Large language models sensitivity to the order of options in multiple-choice questions. In Duh, K., G \' o mez - Adorno, H., and Bethard, S. (eds.), Findings of the Association for Computational Linguistics: NAACL 2024, Mexico City, Mexico, Jun...

    2024 doi

  15. [23]

    Phan, L. et al. Humanity's last exam, 2025. URL https://arxiv.org/abs/2501.14249

    2025 arXiv

  16. [24]

    MAUVE: measuring the gap between neural text and human text using divergence frontiers

    Pillutla, K., Swayamdipta, S., Zellers, R., Thickstun, J., Welleck, S., Choi, Y., and Harchaoui, Z. MAUVE: measuring the gap between neural text and human text using divergence frontiers. In Ranzato, M., Beygelzimer, A., Dauphin, Y. N., Liang, P., and Vaughan, J. W. (eds.), Ad...

    2021

  17. [25]

    Hidden persuaders: Llms' political leaning and their influence on voters

    Potter, Y., Lai, S., Kim, J., Evans, J., and Song, D. Hidden persuaders: Llms' political leaning and their influence on voters. In Al - Onaizan, Y., Bansal, M., and Chen, Y. (eds.), Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, EMNLP 2...

    2024

  18. [26]

    Language models are unsupervised multitask learners

    Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., and Sutskever, I. Language models are unsupervised multitask learners. 2019

    2019

  19. [27]

    R., and Nguyen, A

    Rahmanzadehgervi, P., Bolton, L., Taesiri, M. R., and Nguyen, A. T. Vision language models are blind. In Cho, M., Laptev, I., Tran, D., Yao, A., and Zha, H. (eds.), Computer Vision - ACCV 2024 - 17th Asian Conference on Computer Vision, Hanoi, Vietnam, December 8-12, 2024, Pro...

    2024 doi

  20. [28]

    H., Sattigeri, P., Wornell, G

    Shen, M., Das, S., Greenewald, K. H., Sattigeri, P., Wornell, G. W., and Ghosh, S. Thermometer: Towards universal calibration for large language models. In Forty-first International Conference on Machine Learning, ICML 2024, Vienna, Austria, July 21-27, 2024 . OpenReview.net, ...

    2024

  21. [30]

    The woman worked as a babysitter: On biases in language generation

    Sheng, E., Chang, K.-W., Natarajan, P., and Peng, N. The woman worked as a babysitter: On biases in language generation. In Inui, K., Jiang, J., Ng, V., and Wan, X. (eds.), Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th Inter...

    2019 doi

  22. [31]

    what shapes your bias?

    Shin, J., Song, H., Lee, H., Jeong, S., and Park, J. Ask llms directly, "what shapes your bias?": Measuring social bias in large language models. In Ku, L., Martins, A., and Srikumar, V. (eds.), Findings of the Association for Computational Linguistics, ACL 2024, Bangkok, Thai...

    2024 doi

  23. [32]

    Commonsenseqa: A question answering challenge targeting commonsense knowledge

    Talmor, A., Herzig, J., Lourie, N., and Berant, J. Commonsenseqa: A question answering challenge targeting commonsense knowledge. In Burstein, J., Doran, C., and Solorio, T. (eds.), Proceedings of the 2019 Conference of the North American Chapter of the Association for Computa...

    2019 doi

  24. [33]

    Ulmer, D., Gubri, M., Lee, H., Yun, S., and Oh, S. J. Calibrating large language models using their generations only. In Ku, L., Martins, A., and Srikumar, V. (eds.), Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers...

    2024 doi

  25. [34]

    V., Chi, E

    Wang, X., Wei, J., Schuurmans, D., Le, Q. V., Chi, E. H., Narang, S., Chowdhery, A., and Zhou, D. Self-consistency improves chain of thought reasoning in language models. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, ...

    2023

  26. [35]

    and Ebling, S

    Wu, G. and Ebling, S. Investigating ableism in LLM s through multi-turn conversation. In Dementieva, D., Ignat, O., Jin, Z., Mihalcea, R., Piatti, G., Tetreault, J., Wilson, S., and Zhao, J. (eds.), Proceedings of the Third Workshop on NLP for Positive Impact, pp.\ 202--210, M...

    2024

  27. [36]

    Can llms express their uncertainty? an empirical evaluation of confidence elicitation in llms

    Xiong, M., Hu, Z., Lu, X., Li, Y., Fu, J., He, J., and Hooi, B. Can llms express their uncertainty? an empirical evaluation of confidence elicitation in llms. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024 . Open...

    2024

  28. [37]

    Just rephrase it! uncertainty estimation in closed-source language models via multiple rephrased queries

    Yang, A., Chen, C., and Pitas, K. Just rephrase it! uncertainty estimation in closed-source language models via multiple rephrased queries. CoRR, abs/2405.13907, 2024. doi:10.48550/ARXIV.2405.13907. URL https://doi.org/10.48550/arXiv.2405.13907

    2024 doi arXiv

  29. [38]

    Forcing diffuse distributions out of language models

    Zhang, Y., Schwarzschild, A., Carlini, N., Kolter, Z., and Ippolito, D. Forcing diffuse distributions out of language models. CoRR, abs/2404.10859, 2024. doi:10.48550/ARXIV.2404.10859. URL https://doi.org/10.48550/arXiv.2404.10859

    2024 doi arXiv

  30. [39]

    GPTBIAS: A comprehensive framework for evaluating bias in large language models

    Zhao, J., Fang, M., Pan, S., Yin, W., and Pechenizkiy, M. GPTBIAS: A comprehensive framework for evaluating bias in large language models. CoRR, abs/2312.06315, 2023. doi:10.48550/ARXIV.2312.06315. URL https://doi.org/10.48550/arXiv.2312.06315

    2023 doi arXiv

Pith tools

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