Pith. sign in

REVIEW 4 major objections 6 minor 2 cited by

BiasFilter: An Inference-Time Debiasing Framework for Large Language Models

T0 review · 4 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 BiasFilter claims that social bias in LLM generation can be reduced at inference time by filtering partial outputs with a fairness reward.

desk verdict Sound inference-time debiasing idea with a real judge circularity problem; worth serious refereeing after evaluation is tightened. read the letter →

arxiv 2505.23829 v1 pith:RJZFVMRB submitted 2025-05-28 cs.CL

Xiaoqing Cheng , Ruizhe Chen , Hongying Zan , Yuxiang Jia , Min Peng This is my paper
classification cs.CL
keywords socialbiasmitigationinference-timedebiasingfairnessrewardmodeldirectpreferenceoptimizationopen-endedgenerationlargelanguagemodelsmulti-turndialoguedataset
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

BiasFilter is an attempt to make social-bias mitigation a plug-in operation rather than a retraining procedure. The framework stops a base LLM every fixed number of tokens, scores each partial continuation with a fairness reward, keeps a small set of high-reward candidates, and continues only those candidates. The reward signal comes from an implicit reward model trained with direct preference optimization on a newly constructed fairness preference dataset of biased and unbiased response pairs. On the CEB and FairMT benchmarks the paper reports lower bias scores and bias rates across age, gender, race, and religion for seven open-source models and two API-based models, while perplexity and distinct-2 measures stay roughly the same or improve. If these results hold, debiasing can be added to any generator at inference time without touching its weights.

What carries the argument

The load-bearing mechanism is a beam-search-style filter driven by a partial fairness reward. At each segment of length $l$, every surviving candidate spawns $N_{\mathrm{children}}$ new continuations sampled from the base model; all candidates are scored by the partial reward $r_{\mathrm{partial}}(y_{:K}\mid x) = \sum_{k=0}^{K-1} w_k \log \pi(y_k \mid x, y_{:k}) / \pi_{\mathrm{ref}}(y_k \mid x, y_{:k})$ with weight $w_k = 1/|y_{:k}|$, and the top $N$ survive to the next round. This partial reward is the DPO-style policy ratio of a fairness-trained model $\pi$ versus its reference $\pi_{\mathrm{ref}}$, so it is the trained component that carries the bias judgment. The base generator only samples text; all debiasing work is done by pruning on this reward.

What would settle it

On a held-out set of CEB continuation prompts, run BiasFilter with an unseen base model, log the partial-reward score at every checkpoint, and have independent judges rate the final responses for bias. If a substantial fraction of outputs that passed every checkpoint are rated biased—or if discarded prefixes would have led to fair completions—then the partial reward is not a faithful fairness signal.

Watch

Extended reading notes

Core claim

The paper's central claim is that a sequence does not have to be completed before its fairness is judged. BiasFilter generates in segments, and after each segment it uses a token-level fairness reward to delete the candidates most likely to end in stereotyped text. The authors state that this procedure substantially reduces social bias on age, gender, race, and religion in both continuation and conversation/multi-turn tasks, and that it consistently beats six existing debiasing methods while preserving—or in several cases slightly improving—fluency and diversity. The base generator is never modified; the framework only decides which sampled continuations survive.

Load-bearing premise

The entire filter rests on the premise that the weighted sum of token-level preference ratios for an incomplete text correctly orders continuations by how biased the finished text will be, even for base models the reward model was not trained on.

Editorial extensions

If this is right

  • A single fairness reward model can be reused across many base generators, because BiasFilter never modifies the generator.
  • API-based black-box models can be debiased without any access to weights or gradients.
  • Users can dial latency against bias strength by changing the segment length $l$ and the candidate count $N_{\mathrm{children}}$.
  • Multi-turn fairness improves alongside single-turn fairness, suggesting the filter does not disrupt dialogue structure.
  • Fluency and diversity survive filtering, since reported perplexity and Distinct-2 values are comparable or better after debiasing.

Reading between the lines

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

  • The same filtering loop could in principle carry other safety objectives—toxicity, privacy, or hallucination avoidance—if a token-level reward model for that objective is trained.
  • The design hides a testable premise: partial-sequence reward must match full-sequence bias judgments. A calibration study on unseen base-model families would settle it.
  • Because the filter optimizes the reward rather than the truth of the text, it may be vulnerable to reward hacking—continuations that game the fairness score while remaining biased in ways the judge does not catch.
  • The most practical long-run shape of this idea is a centralized fairness service that wraps arbitrary third-party generators; whether that works depends on how well the reward model's judgments transfer across domains and languages.
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

4 major / 6 minor

Summary. BiasFilter is an inference-time debiasing framework. The authors construct a fairness preference dataset by sampling completions from five LLMs for HolisticBias-derived prompts, scoring them with GPT-4 for stereotypical bias, and human-validating preference pairs. They train a DPO-based reward model (Llama-3-8b-SFT2) with a token-level partial reward defined as a weighted sum of log-ratios. At decoding time, the method maintains a beam of candidate sequences, generates segments of length l, scores each partial sequence with the reward model, prunes to the top-N, and finally returns the highest-scoring complete candidate. Experiments on CEB and FairMT across open-source (Llama, Mistral, Qwen) and API models (GPT-3.5-Turbo, GPT-4o) are reported against six baselines, using GPT-4-based Bias Score/Bias Rate and Regard score, plus perplexity and distinct-2 for quality. The authors claim substantial bias reduction and consistent outperformance of baselines while preserving generation quality.

Significance. The paper addresses a practically important problem: reducing social bias in LLM generation without retraining the base model. The main contributions—a fairness preference dataset, a token-level reward model, and a model-agnostic inference-time filtering algorithm—are well motivated, and the experimental scope is broad (multiple model families, two benchmarks, six baselines, ablations). If the results hold under an independent evaluation instrument, BiasFilter would be a useful addition to the debiasing toolbox, particularly for API-based models. The framework ships a concrete algorithm (Algorithm 1), and the inference-time design leaves the base model untouched. However, the current evidence is weakened by a circularity between the reward-model training signal and the primary evaluation metric, by the absence of error bars or significance tests, and by results that contradict the 'consistently outperforms' claim. These issues are addressable with additional experiments and more careful reporting.

major comments (4)
  1. [§3.1 and §4.1] The primary evaluation metric is the same GPT-4 rubric used to create the reward model's training labels. In §3.1, GPT-4 scores responses with the prompt in Figure 7 to construct preference pairs, and the DPO reward model in §3.2 is trained on those pairs. In §4.1, the CEB Bias Score is 'computed by GPT-4 based on the degree of stereotypical bias, with the prompt shown in Figure 7.' Algorithm 1 then filters and selects outputs using that reward model. Tables 1, 2, and 4 therefore largely report how well BiasFilter optimizes the exact judge that labeled its training data. The Regard score is an independent metric, but it measures sentiment/favorability rather than stereotype content, so it does not break the loop. I request an independent evaluation of the stored outputs—for example, human annotations on a sample of final generations, or a bias classifier that does not derive from GPT-4 with the same rubric—to verify that the reported debiasing transfers outside the training loop. Without such evidence, the central claim that BiasFilter 'mitigates social bias' is not yet established.
  2. [§4.2, Table 1 and Table 4] The claim that BiasFilter 'consistently outperforms all baselines' is contradicted by the reported numbers. For Meta-Llama-3-8b-Instruct on CEB-conversation, RLRF achieves lower Bias Scores on age (16.5 vs. 19.1) and gender (9.5 vs. 9.8); ARGS achieves lower on age (17.7 vs. 19.1) and gender (8.6 vs. 9.8); SD-Ex achieves lower on age (16.7 vs. 19.1). In Table 4, applying BiasFilter to Qwen2.5-3B-Instruct increases the FairMT FixF Bias Rate from 97.6 (base) to 100.0 (with BiasFilter). The narrative should be adjusted to 'most settings' with a detailed account of where baselines win, or significance tests should be provided to show that the remaining differences are not noise. As written, the paper's headline claim overstates the results.
  3. [§3.2, Eq. (2)] Equation (2) defines the partial reward as a weighted sum of token-level DPO log-ratios, and Algorithm 1 relies on this score to prune candidates at every segment. The paper does not validate that this cumulative score correlates with full-sequence bias, nor that the correlation holds for base models other than Llama-3-8b-SFT2 (the reward model's own family). Because the reward model is a policy trained on full-sequence preferences, its per-token likelihood ratios on partial sequences may not track the fairness of the partial output; if early-segment scores are unreliable, the beam search can discard fair continuations or keep biased ones, and the debiasing effect would not transfer across the diverse base models in Tables 1 and 4. I request a validation experiment that compares partial-reward rankings with full-sequence bias judgments (e.g., GPT-4 or human scores) on held-out prompts for several base models. Without this, the core filtering mechanism is unsubstantiated for the generality claimed.
  4. [§4.1 and Fig. 4] The main hyperparameters are tuned on the evaluation benchmark. The segment length l is fixed to 128 because Fig. 4 shows that this is the best trade-off on CEB-continuation with Meta-Llama-3-8B-Instruct, and §4.1 then uses l=128 for all CEB models and tasks (and l=50 for FairMT, presumably chosen similarly). Since CEB is also the benchmark on which the main results in Table 1 are reported, this is effectively test-set tuning. Additionally, no standard deviations or significance tests are reported anywhere; single-run differences such as 19.1 vs. 17.7 may be within sampling noise. Please either use a development split for hyperparameter selection, or report variance over multiple seeds, or restrict the claim of 'significant' improvements to differences that survive a significance test.
minor comments (6)
  1. [Algorithm 1] Lines 8–10 use the loop variable `yk` both for a sampled continuation and for the prefix that produced it, and the final selected output `y*` is not defined in the algorithm's notation; rewriting with distinct variable names would improve clarity.
  2. [Eq. (2)] The notation `rpartial(y:K | x)` mixes `y:K` and `y:k`; I recommend defining `y_{1:K}` and `y_{1:k}` explicitly and stating that K is the number of tokens in the partial sequence.
  3. [Fig. 4] The sentence "Points closer to the top-right corner indicate a better trade-off" appears to be a typo, because the x-axis is time and the y-axis is Bias Score, so lower-left is the favorable region.
  4. [Appendix B.1] The appendix refers to the tuning procedure as "FairDPO," but the main text calls the model a fairness reward model; please align the names to avoid confusion.
  5. [References] The Mistral-7B model is cited as (Chaplot, 2023); the correct reference is Jiang et al. (2023), and the current citation does not point to a proper Mistral technical report.
  6. [Table 4] Several cells show that BiasFilter worsens the Bias Rate relative to the base model (e.g., Llama-3-8b-Base IntM: 93.3 to 96.4; Qwen2.5-3B FixF: 97.6 to 100.0), yet the text states that the method "consistently improves" fairness across models; these counterexamples should be acknowledged and discussed.

Circularity Check

1 steps flagged · score 6.0 of 10

CEB debiasing gains are partly circular: the DPO reward model is trained on GPT-4 stereotype scores from the same 0-99 prompt used to compute the primary CEB Bias Score, so filtering optimizes the judge that later certifies fairness.

  1. fitted input called prediction [Section 3.1 (dataset construction), Section 3.2 Eq. (1), Section 3.3 Eqs. (3)-(5) and Algorithm 1, Section 4.1 (CEB metric), Appendix A.2 / Figure 7]
    "After generating 30,000 model completions from 6,000 prompts, we use GPT-4 with the prompt (provided in Appendix A.2) to score the level of stereotypical bias in each response. ... The response with the highest bias score is taken as the negative example, and the lowest as the positive, forming our initial fairness preference dataset. ... For CEB, we adopt Bias Score as the primary metric, computed by GPT-4 based on the degree of stereotypical bias, with the prompt shown in Figure 7."

    The Fairness Preference Dataset labels are produced by GPT-4 using the 0-99 stereotypical-bias prompt in Appendix A.2 (shown in Figure 7). The DPO reward model of Eq. (1) is trained on these labels, and Algorithm 1 / Eqs. (3)-(5) select and retain candidates by maximizing that reward model. The paper's primary CEB metric, Bias Score, is then computed by GPT-4 'based on the degree of stereotypical bias, with the prompt shown in Figure 7' — the same prompt that generated the training labels. Thus the reported CEB debiasing improvements measure how well BiasFilter optimizes the exact judge used to construct its reward signal.

full rationale

The main circularity is in the evaluation loop for CEB. The reward model is trained on preference pairs derived from GPT-4's bias scores using the same prompt that defines the primary CEB Bias Score metric. Since the inference-time filtering explicitly maximizes this reward model, the CEB results are partly a measure of how well the method optimizes the judge that labeled its training data. This matches the fitted-input-called-prediction pattern: a parameter (the reward model) is fitted to a scoring rubric, and the headline result is then reported under that same rubric. I did not find self-citation load-bearing circularity: the cited prior work by the authors (e.g., FairMT, BiasGuard, related debiasing methods) is used as benchmark or related work, not to justify a result by assertion. There is also no imported uniqueness theorem or ansatz-smuggling pattern. The Regard score, FairMT bias rate, perplexity/diversity checks, and human validation of the preference pairs are genuinely external or partially independent, which keeps the overall circularity score at 6 rather than higher. However, the central quantitative claim on CEB — the paper's primary benchmark — is not self-contained against an independent judge, so the score is elevated.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The method depends on the selected hyperparameters, the validity of GPT-4-based bias scoring for both training and evaluation, and the cross-model generalizability of the DPO reward model. No new theoretical entities are introduced.

free parameters (5)
  • segment length l = 128 tokens (CEB), 50 tokens (FairMT)
    Chosen based on ablation in Section 4.3 (Figure 4) to balance bias score and time on the CEB-continuation set.
  • Nchildren = 8 (CEB), 6 (FairMT)
    Number of candidate continuations per active segment; the ablation in Figure 5 shows limited sensitivity, but values are hand-picked.
  • beam width N = 4 (CEB), 3 (FairMT)
    Number of top partial sequences retained after each filtering step.
  • DPO beta = not specified in text
    Controls the strength of the implicit reward; default likely from DPO implementation, not reported.
  • LoRA rank / alpha = 16 / 32
    Training hyperparameters for the reward model in Appendix B.1.
assumptions (4)
  • domain assumption GPT-4 bias scores are a valid measure of social bias.
    Used to construct the training pairs (Section 3.1) and to score the CEB evaluation (Section 4.1).
  • domain assumption The DPO implicit reward decomposes into a weighted sum of token-level log-ratios (Eq. 2).
    This is a design choice, not a proven equivalence; the paper follows Qiu et al. without validation on fairness.
  • domain assumption The fairness preference dataset sampled from HolisticBias is representative of real-world biased prompts.
    The dataset is based on a subset of templates and four social groups; coverage may be limited.
  • domain assumption The reward model trained on Llama-3-8b-SFT2 generalizes to other base models and API models.
    The paper assumes the same reward model can score outputs of any base model without recalibration, which is supported only by experiments, not by theory.

how reviews work

0 comments
Cite this review

Pith. "Pith review of BiasFilter: An Inference-Time Debiasing Framework for Large Language Models." pith.science (2026). https://pith.science/paper/RJZFVMRB

@misc{pith2026250523829,
  author       = {Pith},
  title        = {Pith review of: BiasFilter: An Inference-Time Debiasing Framework for Large Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RJZFVMRB}},
  note         = {Machine review of arXiv:2505.23829}
}
read the original abstract

Mitigating social bias in large language models (LLMs) has become an increasingly important research objective. However, existing debiasing methods often incur high human and computational costs, exhibit limited effectiveness, and struggle to scale to larger models and open-ended generation tasks. To address these limitations, this paper proposes BiasFilter, a model-agnostic, inference-time debiasing framework that integrates seamlessly with both open-source and API-based LLMs. Instead of relying on retraining with balanced data or modifying model parameters, BiasFilter enforces fairness by filtering generation outputs in real time. Specifically, it periodically evaluates intermediate outputs every few tokens, maintains an active set of candidate continuations, and incrementally completes generation by discarding low-reward segments based on a fairness reward signal. To support this process, we construct a fairness preference dataset and train an implicit reward model to assess token-level fairness in generated responses. Extensive experiments demonstrate that BiasFilter effectively mitigates social bias across a range of LLMs while preserving overall generation quality.

Figures

Figures reproduced from arXiv: 2505.23829 by the authors.

Figure 1
Figure 1. Overview of BIASFILTER. BIASFILTER can be seamlessly integrated with a base model (e.g., exist￾ing LLMs or APIs) to periodically assess the fairness of generated outputs. Generations that fail the fairness check are filtered out, while those that pass are allowed to proceed in the generation process. communities, reinforce societal stereotypes, and undermine fairness (Bender et al., 2021; Birhane and Prabhu, 2021), … view at source ↗
Figure 2
Figure 2. Illustration of the dataset construction process. We construct the Fairness Preference Dataset by sampling responses from multiple base models and obtaining high-quality annotations from multi-round GPT-4 and human validation. Detailed statistics regarding the tasks and bias attributes of the dataset are shown on the right. models (LLMs) or APIs. Specifically, we intro￾duce an auxiliary implicit reward model to eval… view at source ↗
Figure 3
Figure 3. Illustration of the BIASFILTER Framework. BIASFILTER mitigates bias by employing a fairness reward model to evaluate the fairness of intermediate generations from the base model, filtering out low-reward candidates, and retaining fair ones for continued generation. Without requiring any modification to the base model, this process promotes final outputs that are both high-quality and unbiased. Algorithm 1 BIASFILTER… view at source ↗
Figures from the paper (5 more)
Figure 5
Figure 5. Figure 5: Effect of sample size Nchildren. Bias Scores are compared for Nchildren from 3 to 7. Age Gender Race Religion Social Groups 0.0 2.5 5.0 7.5 10.0 12.5 15.0 17.5 Bias Score Base Greedy Stochastic Top-p [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 7
Figure 7. Figure 7: Prompt for Evaluation. cessible and proprietary large language models. C.2 Additional Evaluation on FairMT We further assess the effectiveness of our method on the FairMT dataset using the Regard score as the evaluation metric. We compare our method with several strong…
Figure 8
Figure 8. Figure 8: Effect of sample size Nchildren. We compare bias scores as Nchildren ranges from 4 to 20. pool enables the reward model (introduced in Sec￾tion 3.2) to better identify fairer candidates, thereby reducing the overall bias in the generated text. D Case Studies To clearly…
Figure 9
Figure 9. Figure 9: Examples from Qwen2.5-14B on CEB-continuation and FairMT. [PITH_FULL_IMAGE:figures/full_fig_p016_9.png]
Figure 12
Figure 12. Figure 12: Examples from Mistral-7b on FairMT [PITH_FULL_IMAGE:figures/full_fig_p018_12.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

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

  1. AI4Research: A Survey of Artificial Intelligence for Scientific Research

    cs.CL 2025-07 conditional novelty 4.0 of 10

    A survey that organizes AI-for-research work into five tasks, comprehension, survey, discovery, writing, and peer review, and compiles associated tools and benchmarks.

  2. Detection, Classification, and Mitigation of Gender Bias in Large Language Models

    cs.CL 2025-06 conditional novelty 4.0 of 10

    A Chinese gender-bias system using SFT, chain-of-thought, and DPO with GPT-4-generated preference pairs reports top validation scores and first place on all three NLPCC 2025 subtasks.

Reference graph

Works this paper leans on

67 extracted references · 16 canonical work pages · cited by 2 Pith papers

  1. [1]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...

  2. [2]

    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 word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, and 1 others. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774

    arXiv 2023

  4. [4]

    AI@Meta. 2024. https://github.com/meta-llama/llama3/blob/main/MODEL_CARD.md Llama 3 model card

    2024

  5. [5]

    Ahmed Allam. 2024. Biasdpo: Mitigating bias in language models through direct preference optimization. arXiv preprint arXiv:2407.13928

    arXiv 2024

  6. [6]

    Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain, Stanislav Fort, Deep Ganguli, Tom Henighan, and 1 others. 2022. Training a helpful and harmless assistant with reinforcement learning from human feedback. arXiv preprint arXiv:2204.05862

    arXiv 2022

  7. [7]

    Emily M Bender, Timnit Gebru, Angelina McMillan-Major, and Shmargaret Shmitchell. 2021. On the dangers of stochastic parrots: Can language models be too big? In Proceedings of the 2021 ACM conference on fairness, accountability, and transparency, pages 610--623

    2021

  8. [8]

    Abeba Birhane and Vinay Uday Prabhu. 2021. Large image datasets: A pyrrhic win for computer vision? In 2021 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 1536--1546. IEEE

    work page 2021

Show all 67 references
  1. [9]

    Devendra Singh Chaplot. 2023. Albert q. jiang, alexandre sablayrolles, arthur mensch, chris bamford, devendra singh chaplot, diego de las casas, florian bressand, gianna lengyel, guillaume lample, lucile saulnier, l \'e lio renard lavaud, marie-anne lachaux, pierre stock, teve...

    2023 arXiv

  2. [10]

    Ruizhe Chen, Wenhao Chai, Zhifei Yang, Xiaotian Zhang, Joey Tianyi Zhou, Tony Quek, Soujanya Poria, and Zuozhu Liu. 2025 a . Diffpo: Diffusion-styled preference optimization for efficient inference-time alignment of large language models. arXiv preprint arXiv:2503.04240

    2025 arXiv

  3. [11]

    Ruizhe Chen, Tianxiang Hu, Yang Feng, and Zuozhu Liu. 2024 a . Learnable privacy neurons localization in language models. arXiv preprint arXiv:2405.10989

    2024 arXiv

  4. [12]

    Ruizhe Chen, Yichen Li, Jianfei Yang, Yang Feng, Joey Tianyi Zhou, Jian Wu, and Zuozhu Liu. 2025 b . Identifying and mitigating social bias knowledge in language models. In Findings of the Association for Computational Linguistics: NAACL 2025, pages 651--672

    2025

  5. [13]

    Ruizhe Chen, Jianfei Yang, Huimin Xiong, Jianhong Bai, Tianxiang Hu, Jin Hao, Yang Feng, Joey Tianyi Zhou, Jian Wu, and Zuozhu Liu. 2023. Fast model debias with machine unlearning. Advances in Neural Information Processing Systems, 36:14516--14539

    2023

  6. [14]

    Ruizhe Chen, Xiaotian Zhang, Meng Luo, Wenhao Chai, and Zuozhu Liu. 2024 b . Pad: Personalized alignment of llms at decoding-time. arXiv preprint arXiv:2410.04070

    2024 arXiv

  7. [15]

    Ruoxi Cheng, Haoxuan Ma, Shuirong Cao, and Tianyu Shi. 2024. Rlrf: Reinforcement learning from reflection through debates as feedback for bias mitigation in llms. arXiv preprint arXiv:2404.10160

    2024 arXiv

  8. [16]

    John Joon Young Chung, Ece Kamar, and Saleema Amershi. 2023. Increasing diversity while maintaining accuracy: Text data generation with large language models and human interventions. arXiv preprint arXiv:2306.04140

    2023 arXiv

  9. [17]

    LLaMA-Factory Contributors. 2023. Llama-factory: Easy and efficient fine-tuning of llama-based models. https://github.com/huggingface/llama-factory

    2023

  10. [18]

    Shizhe Diao, Rui Pan, Hanze Dong, Ka Shun Shum, Jipeng Zhang, Wei Xiong, and Tong Zhang. 2023. Lmflow: An extensible toolkit for finetuning and inference of large foundation models. arXiv preprint arXiv:2306.12420

    2023 arXiv

  11. [19]

    Li Du, Zhouhao Sun, Xiao Ding, Yixuan Ma, Yang Zhao, Kaitao Qiu, Ting Liu, and Bing Qin. 2024. Causal-guided active learning for debiasing large language models. arXiv preprint arXiv:2408.12942

    2024 arXiv

  12. [20]

    Satyam Dwivedi, Sanjukta Ghosh, and Shivam Dwivedi. 2023. Breaking the bias: Gender fairness in llms using prompt engineering and in-context learning. Rupkatha Journal on Interdisciplinary Studies in Humanities, 15(4)

    2023

  13. [21]

    Sana Ebrahimi, Kaiwen Chen, Abolfazl Asudeh, Gautam Das, and Nick Koudas. 2024. Axolotl: fairness through assisted self-debiasing of large language model outputs. arXiv preprint arXiv:2403.00198

    2024 arXiv

  14. [22]

    Jessica Echterhoff, Yao Liu, Abeer Alessa, Julian McAuley, and Zexue He. 2024. Cognitive bias in decision-making with llms. arXiv preprint arXiv:2403.00811

    2024 arXiv

  15. [23]

    Zhiting Fan, Ruizhe Chen, Tianxiang Hu, and Zuozhu Liu. 2024 a . Fairmt-bench: Benchmarking fairness for multi-turn dialogue in conversational llms. arXiv preprint arXiv:2410.19317

    2024 arXiv

  16. [24]

    Zhiting Fan, Ruizhe Chen, and Zuozhu Liu. 2025. Biasguard: A reasoning-enhanced bias detection tool for large language models. arXiv preprint arXiv:2504.21299

    2025 arXiv

  17. [25]

    Zhiting Fan, Ruizhe Chen, Ruiling Xu, and Zuozhu Liu. 2024 b . Biasalert: A plug-and-play tool for social bias detection in llms. arXiv preprint arXiv:2407.10241

    2024 arXiv

  18. [26]

    Isabel O Gallegos, Ryan A Rossi, Joe Barrow, Md Mehrab Tanjim, Sungchul Kim, Franck Dernoncourt, Tong Yu, Ruiyi Zhang, and Nesreen K Ahmed. 2024 a . Bias and fairness in large language models: A survey. Computational Linguistics, 50(3):1097--1179

    2024

  19. [27]

    Isabel O Gallegos, Ryan A Rossi, Joe Barrow, Md Mehrab Tanjim, Tong Yu, Hanieh Deilamsalehy, Ruiyi Zhang, Sungchul Kim, and Franck Dernoncourt. 2024 b . Self-debiasing large language models: Zero-shot recognition and reduction of stereotypes. arXiv preprint arXiv:2402.01981

    2024 arXiv

  20. [28]

    Mingqi Gao, Xinyu Hu, Xunjian Yin, Jie Ruan, Xiao Pu, and Xiaojun Wan. 2025. Llm-based nlg evaluation: Current status and challenges. Computational Linguistics, pages 1--28

    2025

  21. [29]

    Yingqiang Ge, Wenyue Hua, Kai Mei, Juntao Tan, Shuyuan Xu, Zelong Li, Yongfeng Zhang, and 1 others. 2023. Openagi: When llm meets domain experts. Advances in Neural Information Processing Systems, 36:5539--5568

    2023

  22. [30]

    Yufei Guo, Muzhe Guo, Juntao Su, Zhou Yang, Mengqiu Zhu, Hongfei Li, Mengyang Qiu, and Shuo Shuo Liu. 2024. Bias in large language models: Origin, evaluation, and mitigation. arXiv preprint arXiv:2411.10915

    2024 arXiv

  23. [31]

    S Hallinan, A Liu, Y Choi, and M Sap. 2022. Detoxifying text with marco: Controllable revision with experts and anti-experts (2022)

    2022

  24. [32]

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, and 1 others. 2022. Lora: Low-rank adaptation of large language models. ICLR, 1(2):3

    2022

  25. [33]

    Dong Huang, Qingwen Bu, Jie Zhang, Xiaofei Xie, Junjie Chen, and Heming Cui. 2023. Bias testing and mitigation in llm-based code generation. arXiv preprint arXiv:2309.14345

    2023 arXiv

  26. [34]

    Binyuan Hui, Jian Yang, Zeyu Cui, Jiaxi Yang, Dayiheng Liu, Lei Zhang, Tianyu Liu, Jiajun Zhang, Bowen Yu, Keming Lu, and 1 others. 2024. Qwen2. 5-coder technical report. arXiv preprint arXiv:2409.12186

    2024 arXiv

  27. [35]

    Masahiro Kaneko, Danushka Bollegala, Naoaki Okazaki, and Timothy Baldwin. 2024. Evaluating gender bias in large language models via chain-of-thought prompting. arXiv preprint arXiv:2401.15585

    2024 arXiv

  28. [36]

    Maxim Khanov, Jirayu Burapacheep, and Yixuan Li. 2024. Args: Alignment as reward-guided search. arXiv preprint arXiv:2402.01694

    2024 arXiv

  29. [37]

    Minbeom Kim, Hwanhee Lee, Kang Min Yoo, Joonsuk Park, Hwaran Lee, and Kyomin Jung. 2022. Critic-guided decoding for controlled text generation. arXiv preprint arXiv:2212.10938

    2022 arXiv

  30. [38]

    JESSICA B Kuntz and ELISE C Silva. 2023. Who authors the internet. Analyzing Gender Diversity in ChatGPT-3 Training Data. Pitt Cyber: University of Pittsburgh

    2023

  31. [39]

    Harrison Lee, Samrat Phatale, Hassan Mansoor, Thomas Mesnard, Johan Ferret, Kellie Lu, Colton Bishop, Ethan Hall, Victor Carbune, Abhinav Rastogi, and 1 others. 2023. Rlaif vs. rlhf: Scaling reinforcement learning from human feedback with ai feedback. arXiv preprint arXiv:2309.00267

    2023 arXiv

  32. [40]

    Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov, and Luke Zettlemoyer. 2019. Bart: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. arXiv preprint arXiv:1910.13461

    2019 arXiv

  33. [41]

    Jingling Li, Zeyu Tang, Xiaoyu Liu, Peter Spirtes, Kun Zhang, Liu Leqi, and Yang Liu. 2024 a . Steering llms towards unbiased responses: A causality-guided debiasing framework. arXiv preprint arXiv:2403.08743

    2024 arXiv

  34. [42]

    Jiwei Li, Michel Galley, Chris Brockett, Jianfeng Gao, and Bill Dolan. 2015. A diversity-promoting objective function for neural conversation models. arXiv preprint arXiv:1510.03055

    2015 arXiv

  35. [43]

    Yichen Li, Zhiting Fan, Ruizhe Chen, Xiaotang Gai, Luqi Gong, Yan Zhang, and Zuozhu Liu. 2025. Fairsteer: Inference time debiasing for llms with dynamic activation steering. arXiv preprint arXiv:2504.14492

    2025 arXiv

  36. [44]

    Yingji Li, Mengnan Du, Rui Song, Xin Wang, Mingchen Sun, and Ying Wang. 2024 b . Mitigating social biases of pre-trained language models via contrastive self-debiasing with double data augmentation. Artificial Intelligence, 332:104143

    2024

  37. [45]

    Shuo Shuo Liu. 2024. Unified transfer learning in high-dimensional linear regression. In International Conference on Artificial Intelligence and Statistics, pages 1036--1044. PMLR

    2024

  38. [46]

    Xin Liu, Muhammad Khalifa, and Lu Wang. 2023 a . Bolt: Fast energy-based controlled text generation with tunable biases. arXiv preprint arXiv:2305.12018

    2023 arXiv

  39. [47]

    Yang Liu, Dan Iter, Yichong Xu, Shuohang Wang, Ruochen Xu, and Chenguang Zhu. 2023 b . G-eval: Nlg evaluation using gpt-4 with better human alignment. arXiv preprint arXiv:2303.16634

    2023 arXiv

  40. [48]

    Zhongkun Liu, Zheng Chen, Mengqi Zhang, Zhaochun Ren, Zhumin Chen, and Pengjie Ren. 2024. Zero-shot position debiasing for large language models. CoRR

    2024

  41. [49]

    Nicholas Meade, Spandana Gella, Devamanyu Hazarika, Prakhar Gupta, Di Jin, Siva Reddy, Yang Liu, and Dilek Hakkani-T \"u r. 2023. Using in-context learning to improve dialogue safety. arXiv preprint arXiv:2302.00871

    2023 arXiv

  42. [50]

    Roberto Navigli, Simone Conia, and Bj \"o rn Ross. 2023. Biases in large language models: origins, inventory, and discussion. ACM Journal of Data and Information Quality, 15(2):1--21

    2023

  43. [51]

    OpenAI. 2023. https://platform.openai.com/docs/models/gpt-3-5 Gpt-3.5-turbo

    2023

  44. [52]

    Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, and 1 others. 2022. Training language models to follow instructions with human feedback. Advances in neural information processing sys...

    2022

  45. [53]

    Jiahao Qiu, Yifu Lu, Yifan Zeng, Jiacheng Guo, Jiayi Geng, Huazheng Wang, Kaixuan Huang, Yue Wu, and Mengdi Wang. 2024. Treebon: Enhancing inference-time alignment with speculative tree-search and best-of-n sampling. arXiv preprint arXiv:2410.16033

    2024 arXiv

  46. [54]

    Yao Qu and Jue Wang. 2024. Performance and biases of large language models in public opinion simulation. Humanities and Social Sciences Communications, 11(1):1--13

    2024

  47. [55]

    Rafael Rafailov, Joey Hejna, Ryan Park, and Chelsea Finn. 2024. From r to Q^* : Your language model is secretly a q-function. arXiv preprint arXiv:2404.12358

    2024 arXiv

  48. [56]

    Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. 2023. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 36:53728--53741

    2023

  49. [57]

    Danielle Saunders, Rosie Sallis, and Bill Byrne. 2021. First the worst: Finding better gender translations during beam search. arXiv preprint arXiv:2104.07429

    2021 arXiv

  50. [58]

    Standard Self-Reflection. 2024. Self-contrast: Better reflection through inconsistent solving perspectives

    2024

  51. [59]

    i'm sorry to hear that

    Eric Michael Smith, Melissa Hall, Melanie Kambadur, Eleonora Presani, and Adina Williams. 2022. https://doi.org/10.48550/ARXIV.2205.09209 "i'm sorry to hear that": Finding new biases in language models with a holistic descriptor dataset . arXiv preprint

    2022 doi arXiv

  52. [60]

    Schrasing Tong, Eliott Zemour, Rawisara Lohanimit, and Lalana Kagal. 2024. Towards resource efficient and interpretable bias mitigation in large language models. arXiv preprint arXiv:2412.01711

    2024

  53. [61]

    Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, and 1 others. 2023. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288

    2023 arXiv

  54. [62]

    Boxin Wang, Wei Ping, Chaowei Xiao, Peng Xu, Mostofa Patwary, Mohammad Shoeybi, Bo Li, Anima Anandkumar, and Bryan Catanzaro. 2022. Exploring the limits of domain-adaptive training for detoxifying large-scale language models. Advances in Neural Information Processing Systems, ...

    2022

  55. [63]

    Song Wang, Peng Wang, Tong Zhou, Yushun Dong, Zhen Tan, and Jundong Li. 2024. Ceb: Compositional evaluation benchmark for fairness in large language models. arXiv preprint arXiv:2407.02408

    2024 arXiv

  56. [64]

    An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, Guanting Dong, Haoran Wei, Huan Lin, Jialong Tang, Jialin Wang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Ma, and 40 others. 2024. Qwen2 technical repo...

    2024 arXiv

  57. [65]

    Congzhi Zhang, Linhai Zhang, Jialong Wu, Yulan He, and Deyu Zhou. 2024 a . Causal prompting: Debiasing large language model prompting based on front-door adjustment. arXiv preprint arXiv:2403.02738

    2024 arXiv

  58. [66]

    Wenqi Zhang, Yongliang Shen, Linjuan Wu, Qiuying Peng, Jun Wang, Yueting Zhuang, and Weiming Lu. 2024 b . Self-contrast: Better reflection through inconsistent solving perspectives. arXiv preprint arXiv:2401.02009

    2024 arXiv

  59. [67]

    Xiaotian Zhang, Ruizhe Chen, Yang Feng, and Zuozhu Liu. 2025. Persona-judge: Personalized alignment of large language models via token-level self-judgment. arXiv preprint arXiv:2504.12663

    2025 arXiv

Pith tools

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