Pith. sign in

REVIEW 4 major objections 4 minor 2 cited by

LLM judges are systematically biased, and a new training method measurably reduces it.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-04 05:49 UTC pith:BJGSCNZK

load-bearing objection Useful benchmark, but the debiasing result is likely inflated by an asymmetric metric and the eval/train pipeline sharing bias recipes. the 4 major comments →

arxiv 2603.08091 v2 pith:BJGSCNZK submitted 2026-03-09 cs.CL

Toward Robust LLM-Based Judges: Taxonomic Bias Evaluation and Debiasing Optimization

Hongli Zhou , Hui Huang , Rui Zhang , Kehai Chen , Bing Xu , Conghui Zhu , Tiejun Zhao , Muyun Yang This is my paper
classification cs.CL
keywords judgment biasLLM-as-a-judgeautomated evaluationreward modelingbias taxonomydebiasingGRPOcontrastive learning
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper argues that large language models used as judges — the systems that score and rank AI-generated answers — are systematically swayed by task-irrelevant cues such as verbosity, formatting, confidence, majority opinion, and identity. To make that argument testable, it introduces JudgeBiasBench, a benchmark with a four-part taxonomy and twelve concrete bias types, plus a metric called Bias Sensitivity Rate that isolates bias from ordinary reasoning errors. Across many generative and discriminative judges, the paper shows that bias is widespread, persists in strong models, and is not predicted by normal accuracy. It then proposes bias-aware training — reinforcement learning for generative judges and contrastive learning for discriminative judges — and reports that it sharply lowers bias sensitivity while keeping general evaluation quality roughly unchanged. A sympathetic reader would care because these judges are also reward models in alignment pipelines, so unmeasured bias can become reward hacking.

Core claim

On the paper's own terms, the discovery is that judgment bias in LLM-based judges is a measurable, general phenomenon under a clear definition: a judge is biased when changing a task-irrelevant attribute — style, context, layout, or identity — flips an otherwise correct preference. JudgeBiasBench operationalizes this with controlled injections over twelve bias types and a consistency filter that keeps only instances where the original quality ordering is preserved. Using the resulting measure, the authors find that most judges, including strong reasoning models and high-accuracy reward models, reverse correct judgments at rates that vary from roughly 10% to over 40%, and that length, positio

What carries the argument

The machinery is JudgeBiasBench: a controlled bias-injection pipeline that rewrites or recontextualizes preference pairs so that the task-relevant quality stays fixed while one task-irrelevant factor varies, then uses a strong verifier to filter out pairs whose preference would actually reverse. The quantitative instrument is the Bias Sensitivity Rate (BSR), defined as the share of originally correct judgments that turn wrong after injection. The debiasing machinery is a set of bias-augmented preference pairs — a chosen response contrasted with several rejected responses that look superficially attractive for different bias reasons — trained with GRPO (a reinforcement-learning objective) for

Load-bearing premise

The entire benchmark rests on the claim that the verifier used to filter bias-injected pairs agrees with true quality; if that verifier shares the target biases, the retained test pairs are mislabeled and every BSR number is systematically wrong.

What would settle it

Take a random sample of JudgeBiasBench's filtered pairs, have independent human annotators relabel which response is better, and check agreement with the verifier's labels. If human agreement is notably lower than the verifier's internal consistency, the benchmark is measuring verifier bias rather than judge bias — and the reported BSR reductions would need re-interpreting.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If BSR is adopted as a reporting standard, judge evaluations will need to report robustness alongside accuracy, because the paper shows high clean accuracy does not imply low bias.
  • Bias-aware training can cut BSR substantially on the twelve tested bias types with only a small change in general benchmark agreement, making it a practical recipe for reward-model training.
  • Because length, position, and beauty biases persist even in strong models, any evaluation pipeline should treat these as residual known risks.
  • The taxonomy's separation of bias from judgment errors gives downstream users a cleaner diagnostic: when a judge fails, one can ask whether it is a reasoning failure or a systematic preference for a task-irrelevant cue.
  • Discriminative judges' elevated gender and race biases imply alignment pipelines using them need explicit fairness checks.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The benchmark's numbers depend on the verifier's own impartiality; a natural extension is to re-label the filtered pairs with human annotators or with several independent verifiers and compare BSR estimates.
  • The bias-aware training recipe could be tested on bias types outside its taxonomy, such as language, dialect, or formatting quirks, since the contrastive and reinforcement objectives are not tied to the twelve predefined cues.
  • If bias-aware judges are used inside RLHF, one would expect fewer spurious reward signals and less reward hacking; that is a testable downstream prediction the paper does not run.
  • The tension between bias-aware data proportion and general performance suggests an optimal-calibration curve; a practical extension would be a scheduling rule that anneals bias-aware data as training progresses.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper proposes JudgeBiasBench, a benchmark for evaluating judgment biases in LLM-based judges, with a taxonomy of 12 bias types across 4 dimensions (superficial quality, context, presentation, diversity). Bias-augmented instances are created through counterfactual rewriting, context distraction, layout manipulation, and identity injection; a verifier model filters instances to preserve the original preference. The paper evaluates a variety of generative and discriminative judges on JudgeBiasBench using a new metric, Bias Sensitivity Rate (BSR), and proposes a bias-aware training framework: SFT plus GRPO for generative judges, and InfoNCE contrastive learning for discriminative judges. The central claim is that this training substantially reduces judgment bias while largely preserving general evaluation capability on RewardBench, JudgeBench, RMB, and RM-Bench.

Significance. The problem is timely and important: LLM-as-a-judge is widely deployed, and a unified taxonomic benchmark covering both generative and discriminative judges would be a useful community resource. The paper's empirical coverage is substantial (12 bias types, 18 judges), and the debiasing recipe is clearly described. However, the current evidence does not yet support the central claim. The BSR metric is asymmetric and can be lowered by replacing one bias with another; the training and test bias-injection pipelines are nearly identical; and the ground-truth verification rests on unvalidated proprietary verifiers. These issues are fixable but require additional analysis, so the paper is not acceptable in its present form.

major comments (4)
  1. [Eq. (4), Table 5] The BSR metric is asymmetric: it counts only flips from originally correct to incorrect and ignores flips from originally incorrect to correct. A judge with a systematic opposite bias (e.g., a first-position preference) will have low BSR because its original errors are excluded from the denominator. This is not a theoretical edge case: in Table 5, every bias-aware row shows Acc_inj > Acc_ori (generative Qwen2.5: 77.4 vs 72.1; discriminative Qwen2.5: 80.5 vs 75.5; Llama-3.1: 83.7 vs 77.2). Since the benchmark places the chosen response first for most bias types (Section 3.3.2), this pattern is exactly what a first-position heuristic would produce. The reported BSR reductions are therefore not by themselves evidence of bias removal. Please report a symmetric metric (e.g., total flip rate or both directions of the confusion matrix) and show that Acc_inj is not inflated by a position heurist
  2. [Sections 3.3.2 and 5.1] The training data for debiasing uses the same bias-injection recipe as the test benchmark. Section 5.1 stage 2 explicitly says context and presentation biases are introduced by the manipulation methods of Section 3.3.2, and the same rewriting/injection templates are used for the other categories. The low BSR of the bias-aware models on JudgeBiasBench may therefore reflect overfitting to the specific injection templates rather than a general reduction in bias. The abstract and Section 5.4.2 claim robustness, but the evaluation is on the same distribution. Please add a held-out test set with bias types or injection styles not seen in training, or at least report per-bias-type results that separate seen and unseen manipulations.
  3. [Section 3.3.3 and Section 5.1 stage 3] The validity of JudgeBiasBench rests on the assumption that Gemini-2.5-Pro (test) and GPT-4o (training) reliably verify that the original preference is preserved after bias injection. No human validation, inter-verifier agreement, or error analysis of the verifier is reported. If the verifier shares the target bias or imposes its own preference, the filtered instances are mislabeled and every BSR number is systematically wrong. Please provide a human-evaluation sample of the filtered test set and report per-bias-type verifier agreement; a robust approach would use multiple verifiers and retain only instances with consensus.
  4. [Tables 4-6] No confidence intervals, bootstrap estimates, or significance tests are reported. Sample sizes per bias type are 291–500, and differences such as BSR 10.8 vs 20.7 in Table 5 may be within noise for some columns. The claim of a 'substantial reduction' in judgment bias should be accompanied by uncertainty quantification or a significance test. This is a standard requirement for benchmark comparisons and is needed to support the paper's main quantitative conclusion.
minor comments (4)
  1. [Figure 3] The InfoNCE loss is labeled 'InfoNSE' in the figure; correct the typo.
  2. [Section 4.3 Finding 1] The model is referred to as 'Claude-3-7-Sonnet' in Finding 1, but Section 4.1 and Table 4 use 'Claude-3.7-Sonnet'. Unify the name.
  3. [Section 4.2] The definition of BSR in Eq. (4) should explicitly state that flips from incorrect to correct are not counted. Consider adding a sentence and perhaps a full confusion-matrix table.
  4. [Throughout] The paper does not include a data/code availability statement. A clear statement that the benchmark, the bias-aware data, and code will be publicly released would strengthen reproducibility.

Circularity Check

2 steps flagged

Debiasing claim is evaluated on a test set generated by the same perturbation recipe used for its training data, and the BSR metric counts only one direction of preference flips.

specific steps
  1. fitted input called prediction [Section 5.1 (Bias-aware data construction) vs Section 3.3.2 and Section 5.4.1]
    "For context and presentation biases, bias attributes are introduced by manipulation methods as detailed in Section 3.3.2. ... we primarily evaluate our models on JudgeBiasBench."

    The bias-aware training negatives for context and presentation biases are produced with exactly the same layout/context manipulations (position swap, distraction cues) that define the JudgeBiasBench test cases. The generative training reward (Eqs. 7-9) optimizes accuracy on such bias-augmented pairs, and BSR (Eq. 4) is directly a function of accuracy on bias-injected test instances. The reported BSR drop therefore partly reflects fitting the same perturbation recipe, rather than a demonstrated general debiasing ability.

  2. self definitional [Section 4.2, Eq. 4]
    "BSR measures the proportion of originally correct judgments that become incorrect after bias injection. A higher BSR indicates stronger susceptibility to bias-related features ... Conversely, a lower BSR implies greater robustness."

    Eq. 4's numerator counts only correct-to-incorrect flips; the opposite flip (originally incorrect becoming correct after bias injection) is ignored. Because each bias-injected instance in Section 3.3.2 is built to make the rejected response more attractive, a judge with an inverted bias (e.g., always preferring the first/position-favored response) will not flip and will receive a low BSR while remaining fully bias-driven. Table 5's Acc_inj > Acc_ori pattern is consistent with such an acquired inverse heuristic. Thus 'low BSR = unbiased' is a property of the metric's definition, not an established empirical fact.

full rationale

The paper's bias-aware training is not a pure tautology: the training and test sets are drawn from different preference corpora (GRAM vs HelpSteer3), use different generators/verifiers (GPT-4o vs Gemini-2.0-Flash / Gemini-2.5-Pro), and general evaluation capability is checked on four independent external benchmarks (RewardBench, JudgeBench, RMB, RM-Bench). However, the central debiasing claim is measured on JudgeBiasBench, whose context/presentation perturbations are explicitly reused from the training-data construction pipeline, and the BSR metric counts only one cell of the original-vs-injected confusion matrix. These design choices make the headline 'bias-aware training effectively reduces judgment biases' partly a consequence of evaluation design rather than an independent demonstration. The remaining concerns—verifier reliability and generalization to unseen bias attacks—are validity risks rather than circularity, so the overall score is moderate.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The central claims rest on two families of assumptions: (1) that bias can be isolated from quality via LLM rewriting and verifier filtering, and (2) that the training objective evaluated on JudgeBiasBench is a valid proxy for generalized debiasing. The free parameters are mostly un-reported training hyperparameters plus hand-tuned data-mix ratios, so the exact numbers in the headline comparison are not reproducibly pinned down.

free parameters (5)
  • Bias-aware data ratio (generative) = 1:4 (bias-aware : original preference pairs)
    Selected from Table 6 as the operating point that balances low BSR against general benchmark performance; the choice affects the reported improvement.
  • Bias-aware data ratio (discriminative) = 1:1 (bias-augmented rejected responses : original rejected responses)
    Also tuned via Table 6; 4:1 degrades general benchmarks, so the final claim relies on this hand-picked ratio.
  • InfoNCE temperature tau = not reported
    Appears in Eq. (10); no value is given, yet it scales all contrastive scores and affects training dynamics.
  • GRPO KL coefficient beta = not reported
    Appears in Eq. (6); no value is given, yet it controls deviation from the reference policy and thus the debiasing strength.
  • Format reward penalty = -0.5
    Hand-chosen constant in Eq. (8); no ablation or sensitivity analysis is provided.
axioms (5)
  • domain assumption Task-related quality Q(I,R_i) and task-irrelevant attributes Z_i are separable and identifiable (Eqs. 2-3).
    The whole definition of bias depends on the claim that quality and style/context factors can be separated, which is not proven and is itself contested in LLM evaluation research.
  • domain assumption Bias injection alters only the target attribute and preserves the true quality ordering.
    Section 3.3.2 assumes counterfactual rewriting, context distraction, layout manipulation, and identity injection do not change task-relevant correctness; the consistency filter is designed to catch violations but cannot prove none remain.
  • domain assumption Gemini-2.5-Pro is a reliable verifier for consistency filtering.
    Section 3.3.3 keeps only instances where Gemini-2.5-Pro agrees with the original human preference. If this verifier is itself biased, the benchmark labels are corrupted.
  • domain assumption GPT-4o-generated bias-conditioned rejected responses preserve quality ordering after verification.
    Section 5.1 stages 2-3 use GPT-4o to generate and verify 'biased' rejected responses; the training signal inherits any verifier or generator bias.
  • domain assumption Human preferences in HelpSteer3 and GRAM are valid ground truth for quality.
    The benchmark and training data inherit the assumptions, noise, and crowd artifacts of the base preference corpora, which are not independently audited here.

pith-pipeline@v1.3.0-alltime-deepseek · 18946 in / 11759 out tokens · 120617 ms · 2026-08-04T05:49:31.558409+00:00 · methodology

0 comments
read the original abstract

Large language model (LLM)-based judges are widely adopted for automated evaluation and reward modeling, yet their judgments are often affected by judgment biases. Accurately evaluating these biases is essential for ensuring the reliability of LLM-based judges. However, existing studies typically investigate limited biases under a single judge formulation, either generative or discriminative, lacking a comprehensive evaluation. To bridge this gap, we propose JudgeBiasBench, a benchmark for systematically quantifying biases in LLM-based judges. JudgeBiasBench defines a taxonomy of judgment biases across 4 dimensions, and constructs bias-augmented evaluation instances through a controlled bias injection pipeline, covering 12 representative bias types. We conduct extensive experiments across both generative and discriminative judges, revealing that current judges exhibit significant and diverse bias patterns that often compromise the reliability of automated evaluation. To mitigate judgment bias, we propose bias-aware training that explicitly incorporates bias-related attributes into the training process, encouraging judges to disentangle task-relevant quality from bias-correlated cues. By adopting reinforcement learning for generative judges and contrastive learning for discriminative judges, our methods effectively reduce judgment biases while largely preserving general evaluation capability.

Figures

Figures reproduced from arXiv: 2603.08091 by Bing Xu, Conghui Zhu, Hongli Zhou, Hui Huang, Kehai Chen, Muyun Yang, Rui Zhang, Tiejun Zhao.

Figure 1
Figure 1. Figure 1: Two examples of judgment errors. In both cases, the judge may incorrectly [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the JudgeBiasBench construction pipeline. Starting from a pool [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The illustration of our proposed framework. We first construct a bias-augmented [PITH_FULL_IMAGE:figures/full_fig_p018_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Scaling effects of generative judges. 0 20 40 60 80 100 0.5 1.5 3 7 14 Score Model Size (B) Acc_ori Acc_inj BSR 0 20 40 60 80 100 0.5 1.5 3 7 14 Agreement Model Size (B) RewardBench JudgeBench RMB RM-Bench [PITH_FULL_IMAGE:figures/full_fig_p024_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Scaling effects of discriminative judges. [PITH_FULL_IMAGE:figures/full_fig_p024_5.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Evaluating Second-Order Bias of LLMs Through Epistemic Entitlement

    cs.CL 2026-06 unverdicted novelty 7.0

    LLMs display second-order social bias when judging the acceptability of biased texts to demographic groups, as measured by a philosophically derived task that evades guardrails and varies systematically by target group.

  2. Reproducing, Analyzing, and Detecting Reward Hacking in Rubric-Based Reinforcement Learning

    cs.LG 2026-06 unverdicted novelty 7.0

    CHERRL is a new controllable testbed for reproducing, analyzing, and detecting reward hacking in rubric-based RL by injecting known biases into LLM-as-a-Judge systems.

Reference graph

Works this paper leans on

50 extracted references · 1 canonical work pages · cited by 2 Pith papers

  1. [1]

    Chang, X

    Y. Chang, X. Wang, J. Wang, Y. Wu, L. Yang, K. Zhu, H. Chen, X. Yi, C. Wang, Y. Wang, W. Ye, Y. Zhang, Y. Chang, P. S. Yu, Q. Yang, X. Xie, A survey on evaluation of large language models, ACM Trans. Intell. Syst. Technol. 15 (3) (Mar. 2024). doi:10.1145/3641289

    doi:10.1145/3641289 2024

  2. [2]

    H. Zhou, H. Huang, Z. Zhao, L. Han, H. Wang, K. Chen, M. Yang, W. Bao, J. Dong, B. Xu, et al., Lost in benchmarks? rethinking large 25 language model benchmarking with item response theory, arXiv preprint arXiv:2505.15055 (2025)

    arXiv 2025

  3. [3]

    Papineni, S

    K. Papineni, S. Roukos, T. Ward, W.-J. Zhu, Bleu: a method for au- tomatic evaluation of machine translation, in: P. Isabelle, E. Char- niak, D. Lin (Eds.), Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, Association for Computa- tional Linguistics, Philadelphia, Pennsylvania, USA, 2002, pp. 311–318. doi:10.3115/10...

    arXiv 2002

  4. [4]

    J. Gu, X. Jiang, Z. Shi, H. Tan, X. Zhai, C. Xu, W. Li, Y. Shen, S. Ma, H. Liu, et al., A survey on llm-as-a-judge, The Innovation (2024)

    2024

  5. [5]

    Zheng, W.-L

    L. Zheng, W.-L. Chiang, Y. Sheng, S. Zhuang, Z. Wu, Y. Zhuang, Z. Lin, Z. Li, D. Li, E. Xing, H. Zhang, J. E. Gonzalez, I. Stoica, Judging llm- as-a-judge with mt-bench and chatbot arena, in: A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, S. Levine (Eds.), Advances in Neu- ral Information Processing Systems, Vol. 36, Curran Associates, Inc., 2023, ...

    2023

  6. [6]

    Zhong, W

    J. Zhong, W. Shen, Y. Li, S. Gao, H. Lu, Y. Chen, Y. Zhang, W. Zhou, J. Gu, L. Zou, A comprehensive survey of reward models: Taxonomy, applications, challenges, and future, arXiv preprint arXiv:2504.12328 (2025)

    Pith/arXiv arXiv 2025

  7. [7]

    H. Zhou, H. Huang, W. Liu, C. Wang, X. Bu, L. Han, F. Song, M. Yang, W. Jiang, H. Cao, et al., Rm-distiller: Exploiting generative llm for reward model distillation, arXiv preprint arXiv:2601.14032 (2026)

    Pith/arXiv arXiv 2026

  8. [8]

    R. Koo, M. Lee, V. Raheja, J. I. Park, Z. M. Kim, D. Kang, Bench- marking cognitive biases in large language models as evaluators, in: L.-W. Ku, A. Martins, V. Srikumar (Eds.), Findings of the Asso- ciation for Computational Linguistics: ACL 2024, Association for Computational Linguistics, Bangkok, Thailand, 2024, pp. 517–545. doi:10.18653/v1/2024.findings-acl.29

    doi:10.18653/v1/2024.findings-acl.29 2024

  9. [9]

    H. Zhou, H. Huang, Y. Long, B. Xu, C. Zhu, H. Cao, M. Yang, T. Zhao, Mitigating the bias of large language model evaluation, in: M. Sun, J. Liang, X. Han, Z. Liu, Y. He (Eds.), Proceedings of the 23rd Chinese National Conference on Computational Linguistics (Volume 1: Main 26 Conference), Chinese Information Processing Society of China, Taiyuan, China, 20...

    2024

  10. [10]

    T. Liu, W. Xiong, J. Ren, L. Chen, J. Wu, R. Joshi, Y. Gao, J. Shen, Z. Qin, T. Yu, D. Sohn, A. Makarova, J. Z. Liu, Y. Liu, B. Piot, A. Ittycheriah, A. Kumar, M. Saleh, Rrm: Robust reward model train- ing mitigates reward hacking, in: Y. Yue, A. Garg, N. Peng, F. Sha, R. Yu (Eds.), International Conference on Learning Representations, Vol. 2025, 2025, pp...

    2025

  11. [11]

    P. Wang, L. Li, L. Chen, Z. Cai, D. Zhu, B. Lin, Y. Cao, L. Kong, Q. Liu, T. Liu, Z. Sui, Large language models are not fair evaluators, in: L.-W. Ku, A. Martins, V. Srikumar (Eds.), Proceedings of the 62nd An- nual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), Association for Computational Linguistics, Bangkok, Thailan...

    doi:10.18653/v1/2024.acl-long.511 2024

  12. [12]

    Kumar, Y

    A. Kumar, Y. He, A. H. Markosyan, B. Chern, I. Arrieta-Ibarra, Detect- ing prefix bias in llm-based reward models, in: Proceedings of the 2025 ACM Conference on Fairness, Accountability, and Transparency, FAccT ’25, Association for Computing Machinery, New York, NY, USA, 2025, p. 3196–3206. doi:10.1145/3715275.3732204

    arXiv 2025

  13. [13]

    G. H. Chen, S. Chen, Z. Liu, F. Jiang, B. Wang, Humans or LLMs as the judge? a study on judgement bias, in: Y. Al-Onaizan, M. Bansal, Y.-N. Chen (Eds.), Proceedings of the 2024 Conference on Empiri- cal Methods in Natural Language Processing, Association for Com- putational Linguistics, Miami, Florida, USA, 2024, pp. 8301–8327. doi:10.18653/v1/2024.emnlp-main.474

    doi:10.18653/v1/2024.emnlp-main.474 2024

  14. [14]

    Q. Wang, Z. Lou, Z. Tang, N. Chen, X. Zhao, W. Zhang, D. Song, B. He, Assessing judging bias in large reasoning models: An empirical study, in: Second Conference on Language Modeling, 2025

    2025

  15. [15]

    C. Ma, E. Zhang, Y. Zhao, W. Liu, Y. Jia, P. Qing, L. Shi, A. Cohan, Y. Yan, S. Vosoughi, Judging with many minds: Do more perspec- tives mean less prejudice? on bias amplification and resistance in multi- agent based LLM-as-judge, in: C. Christodoulopoulos, T. Chakraborty, C. Rose, V. Peng (Eds.), Findings of the Association for Computational Linguistics...

    doi:10.18653/v1/2025.findings- 2025

  16. [16]

    Tripathi, M

    T. Tripathi, M. Wadhwa, G. Durrett, S. Niekum, Pairwise or point- wise? evaluating feedback protocols for bias in LLM-based evaluation, in: Second Conference on Language Modeling, 2025

    2025

  17. [17]

    J. Mire, Z. T. Aysola, D. Chechelnitsky, N. Deas, C. Zerva, M. Sap, Rejected dialects: Biases against African American language in reward models, in: L. Chiruzzo, A. Ritter, L. Wang (Eds.), Findings of the Association for Computational Linguistics: NAACL 2025, Association for Computational Linguistics, Albuquerque, New Mexico, 2025, pp. 7483–7502. doi:10....

    doi:10.18653/v1/2025.findings-naacl.417 2025

  18. [18]

    Z. Zeng, J. Yu, T. Gao, Y. Meng, T. Goyal, D. Chen, Evaluating large language models at evaluating instruction following, in: B. Kim, Y. Yue, S. Chaudhuri, K. Fragkiadaki, M. Khan, Y. Sun (Eds.), International Conference on Learning Representations, Vol. 2024, 2024, pp. 40193– 40219

    2024

  19. [19]

    J. Ye, Y. Wang, Y. Huang, D. Chen, Q. Zhang, N. Moniz, T. Gao, W. Geyer, C. Huang, P.-Y. Chen, N. Chawla, X. Zhang, Justice or prejudice? quantifying biases in llm-as-a-judge, in: Y. Yue, A. Garg, N. Peng, F. Sha, R. Yu (Eds.), International Conference on Learning Representations, Vol. 2025, 2025, pp. 102351–102390

    2025

  20. [20]

    J. Park, S. Jwa, R. Meiying, D. Kim, S. Choi, OffsetBias: Leverag- ing debiased data for tuning evaluators, in: Y. Al-Onaizan, M. Bansal, Y.-N. Chen (Eds.), Findings of the Association for Computational Lin- guistics: EMNLP 2024, Association for Computational Linguistics, Mi- ami, Florida, USA, 2024, pp. 1043–1067. doi:10.18653/v1/2024.findings- emnlp.57

    doi:10.18653/v1/2024.findings- 2024

  21. [21]

    J. Li, S. Sun, W. Yuan, R.-Z. Fan, h. zhao, P. Liu, Generative judge for evaluating alignment, in: B. Kim, Y. Yue, S. Chaudhuri, K. Fragki- adaki, M. Khan, Y. Sun (Eds.), International Conference on Learning Representations, Vol. 2024, 2024, pp. 27547–27574

    2024

  22. [22]

    L. Zhu, X. Wang, X. Wang, Judgelm: Fine-tuned large language models are scalable judges, in: Y. Yue, A. Garg, N. Peng, F. Sha, R. Yu (Eds.), 28 International Conference on Learning Representations, Vol. 2025, 2025, pp. 51257–51296

    2025

  23. [23]

    N. Chen, Z. Hu, Q. Zou, J. Wu, Q. Wang, B. Hooi, B. He, Judgelrm: Large reasoning models as a judge, arXiv preprint arXiv:2504.00050 (2025)

    arXiv 2025

  24. [24]

    Huang, X

    H. Huang, X. Bu, H. Zhou, Y. Qu, J. Liu, M. Yang, B. Xu, T. Zhao, An empirical study of LLM-as-a-judge for LLM evaluation: Fine-tuned judge model is not a general substitute for GPT-4, in: W. Che, J. Nabende, E. Shutova, M. T. Pilehvar (Eds.), Findings of the As- sociation for Computational Linguistics: ACL 2025, Association for Computational Linguistics,...

    doi:10.18653/v1/2025.findings-acl.306 2025

  25. [25]

    C. Y. Liu, L. Zeng, Y. Xiao, J. He, J. Liu, C. Wang, R. Yan, W. Shen, F. Zhang, J. Xu, et al., Skywork-reward-v2: Scaling preference data curation via human-ai synergy, arXiv preprint arXiv:2507.01352 (2025)

    Pith/arXiv arXiv 2025

  26. [26]

    R. Yang, R. Ding, Y. Lin, H. Zhang, T. Zhang, Regularizing hid- den states enables learning generalizable reward model for llms, in: A. Globerson, L. Mackey, D. Belgrave, A. Fan, U. Paquet, J. Tom- czak, C. Zhang (Eds.), Advances in Neural Information Processing Systems, Vol. 37, Curran Associates, Inc., 2024, pp. 62279–62309. doi:10.52202/079017-1990

    work page doi:10.52202/079017-1990 2024

  27. [27]

    Lambert, V

    N. Lambert, V. Pyatkin, J. Morrison, L. Miranda, B. Y. Lin, K. Chandu, N. Dziri, S. Kumar, T. Zick, Y. Choi, N. A. Smith, H. Hajishirzi, RewardBench: Evaluating reward models for language modeling, in: L. Chiruzzo, A. Ritter, L. Wang (Eds.), Findings of the Association for Computational Linguistics: NAACL 2025, Association for Compu- tational Linguistics,...

    doi:10.18653/v1/2025.findings-naacl.96 2025

  28. [28]

    L. Shi, C. Ma, W. Liang, X. Diao, W. Ma, S. Vosoughi, Judging the judges: A systematic study of position bias in LLM-as-a-judge, in: K. Inui, S. Sakti, H. Wang, D. F. Wong, P. Bhattacharyya, B. Baner- jee, A. Ekbal, T. Chakraborty, D. P. Singh (Eds.), Proceedings of the 14th International Joint Conference on Natural Language Processing and the 4th Confere...

    2025

  29. [29]

    Z. Wang, J. Zeng, O. Delalleau, H.-C. Shin, F. Soares, A. Bukharin, E. Evans, Y. Dong, O. Kuchaiev, Helpsteer3-preference: Open human- annotated preference data across diverse tasks and languages, in: The Thirty-ninth Annual Conference on Neural Information Processing Sys- tems Datasets and Benchmarks Track, 2025

    2025

  30. [30]

    Comanici, E

    G. Comanici, E. Bieber, M. Schaekermann, I. Pasupat, N. Sachdeva, I. Dhillon, M. Blistein, O. Ram, D. Zhang, E. Rosen, et al., Gem- ini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities, arXiv preprint arXiv:2507.06261 (2025)

    Pith/arXiv arXiv 2025

  31. [31]

    Y. Feng, S. Wang, Z. Cheng, Y. Wan, D. Chen, Are we on the right way to assessing llm-as-a-judge?, arXiv preprint arXiv:2512.16041 (2025)

    arXiv 2025

  32. [32]

    Malik, V

    S. Malik, V. Pyatkin, S. Land, J. Morrison, N. A. Smith, H. Hajishirzi, N.Lambert, Rewardbench2: Advancingrewardmodelevaluation, arXiv preprint arXiv:2506.01937 (2025)

    Pith/arXiv arXiv 2025

  33. [33]

    D. Guo, D. Yang, H. Zhang, J. Song, P. Wang, Q. Zhu, R. Xu, R. Zhang, S. Ma, X. Bi, et al., Deepseek-r1 incentivizes reasoning in llms through reinforcement learning, Nature 645 (8081) (2025) 633–638

    2025

  34. [34]

    K. Team, Y. Bai, Y. Bao, G. Chen, J. Chen, N. Chen, R. Chen, Y. Chen, Y. Chen, Y. Chen, et al., Kimi k2: Open agentic intelligence, arXiv preprint arXiv:2507.20534 (2025)

    Pith/arXiv arXiv 2025

  35. [35]

    A. Yang, A. Li, B. Yang, B. Zhang, B. Hui, B. Zheng, B. Yu, C. Gao, C. Huang, C. Lv, et al., Qwen3 technical report, arXiv preprint arXiv:2505.09388 (2025)

    Pith/arXiv arXiv 2025

  36. [36]

    A. Yang, B. Yang, B. Zhang, B. Hui, B. Zheng, B. Yu, C. Li, D. Liu, F. Huang, H. Wei, et al., Qwen2.5 technical report, arXiv preprint arXiv:2412.15115 (2024)

    Pith/arXiv arXiv 2024

  37. [37]

    S. Kim, J. Suk, S. Longpre, B. Y. Lin, J. Shin, S. Welleck, G. Neubig, M. Lee, K. Lee, M. Seo, Prometheus 2: An open source language model 30 specialized in evaluating other language models, in: Y. Al-Onaizan, M. Bansal, Y.-N. Chen (Eds.), Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, Association for Computational...

    doi:10.18653/v1/2024.emnlp-main.248 2024

  38. [38]

    Alexandru, A

    A. Alexandru, A. Calvi, H. Broomfield, J. Golden, K. Dai, M. Leys, M. Burger, M. Bartolo, R. Engeler, S. Pisupati, et al., Atla selene mini: A general purpose evaluation model, arXiv preprint arXiv:2501.17195 (2025)

    Pith/arXiv arXiv 2025

  39. [39]

    C. Y. Liu, L. Zeng, J. Liu, R. Yan, J. He, C. Wang, S. Yan, Y. Liu, Y. Zhou, Skywork-reward: Bag of tricks for reward modeling in llms, arXiv preprint arXiv:2410.18451 (2024)

    Pith/arXiv arXiv 2024

  40. [40]

    C. Wang, Y. Gan, Y. Huo, Y. Mu, Q. He, M. Yang, B. Li, T. Xiao, C. Zhang, T. Liu, J. Zhu, GRAM: A generative foundation reward model for reward generalization, in: A. Singh, M. Fazel, D. Hsu, S. Lacoste- Julien, F. Berkenkamp, T. Maharaj, K. Wagstaff, J. Zhu (Eds.), Pro- ceedings of the 42nd International Conference on Machine Learning, Vol. 267 of Procee...

    2025

  41. [41]

    Huang, Y

    H. Huang, Y. He, H. Zhou, R. Zhang, W. Liu, W. Wang, W. Su, B. Zheng, J. Liu, Think-j: Learning to think for generative llm-as-a- judge, arXiv preprint arXiv:2505.14268 (2025)

    arXiv 2025

  42. [42]

    X. Chen, G. Li, Z. Wang, B. Jin, C. Qian, Y. Wang, H. WANG, Y.Zhang, D.Zhang, T.Zhang, H.Tong, H.Ji, RM-r1: Rewardmodeling as reasoning, in: The Fourteenth International Conference on Learning Representations, 2026

    2026

  43. [43]

    Whitehouse, T

    C. Whitehouse, T. Wang, P. Yu, X. Li, J. E. Weston, I. Kulikov, S. Saha, J1: Incentivizing thinking in LLM-as-a-judge via reinforcement learning, in: The Fourteenth International Conference on Learning Representa- tions, 2026

    2026

  44. [44]

    Z. Shao, P. Wang, Q. Zhu, R. Xu, J. Song, X. Bi, H. Zhang, M. Zhang, Y. Li, Y. Wu, et al., Deepseekmath: Pushing the limits of mathematical 31 reasoning in open language models, arXiv preprint arXiv:2402.03300 (2024)

    Pith/arXiv arXiv 2024

  45. [45]

    A. v. d. Oord, Y. Li, O. Vinyals, Representation learning with con- trastive predictive coding, arXiv preprint arXiv:1807.03748 (2018)

    Pith/arXiv arXiv 2018

  46. [46]

    Grattafiori, A

    A. Grattafiori, A. Dubey, A. Jauhri, A. Pandey, A. Kadian, A. Al-Dahle, A. Letman, A. Mathur, A. Schelten, A. Vaughan, et al., The llama 3 herd of models, arXiv preprint arXiv:2407.21783 (2024)

    Pith/arXiv arXiv 2024

  47. [47]

    Shawe-Taylor, N

    J. Shawe-Taylor, N. Cristianini, Kernel methods for pattern analysis, Cambridge university press, 2004

    2004

  48. [48]

    S. Tan, S. Zhuang, K. Montgomery, W. Tang, A. Cuadron, C. Wang, R. Popa, I. Stoica, Judgebench: A benchmark for evaluating llm-based judges, in: Y. Yue, A. Garg, N. Peng, F. Sha, R. Yu (Eds.), International Conference on Learning Representations, Vol. 2025, 2025, pp. 63277– 63303

    2025

  49. [49]

    E. Zhou, G. Zheng, B. Wang, Z. Xi, S. Dou, R. Bao, W. Shen, L. Xiong, J. Fan, Y. Mou, R. Zheng, T. Gui, Q. Zhang, X. Huang, Rmb: Com- prehensively benchmarking reward models in llm alignment, in: Y. Yue, A. Garg, N. Peng, F. Sha, R. Yu (Eds.), International Conference on Learning Representations, Vol. 2025, 2025, pp. 26543–26589

    2025

  50. [50]

    Y. Liu, Z. Yao, R. Min, Y. Cao, L. Hou, J. Li, Rm-bench: Benchmarking reward models of language models with subtlety and style, in: Y. Yue, A. Garg, N. Peng, F. Sha, R. Yu (Eds.), International Conference on Learning Representations, Vol. 2025, 2025, pp. 44323–44355. 32

    2025