Pith. sign in

REVIEW 4 major objections 5 minor 44 references

Fair on the Surface? Benchmarking Hidden-Output Fairness Gaps in LLM Recommenders

T0 review · 4 major / 5 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read Output-level fairness audits of LLM recommenders are structurally incomplete because hidden representation shifts and recommendation shifts decouple.

desk verdict Solid and useful benchmark scaffolding for hidden-vs-output fairness in LLM recommenders, but the central claim of pervasive decoupling is not established by the paper's own metric-robustness ablation. read the letter →

arxiv 2608.08284 v1 pith:FTDLLJLC submitted 2026-08-08 cs.AI

Chan Aristella Lu , Arya Fayyazi , Junhao Zhang , Saeid Shokoufa , Yue Xing , Zhen Xiang , Kyu Hyung Lee , Mehdi Kamal
show 1 more author
Massoud Pedram
This is my paper · ORCID
classification cs.AI
keywords fairnessauditingLLMrecommendershiddenrepresentationscounterfactualprobesrepresentation-outputalignmentactivationsteeringinternalbiasoutput
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 checking only the recommendations an LLM produces cannot establish whether the model is fair, because the hidden representations behind those recommendations can shift in ways the outputs do not reveal. It introduces FairGap, a benchmark that measures observable output shift and internal representation shift from counterfactual identity probes across gender, age, and race, applied to six open-weight LLM families and three recommendation domains. Across conditions the two measures align only weakly, with the Spearman correlation rarely above 0.22, and a non-negligible share of users showing stable outputs alongside substantial internal shifts. The paper further claims this decoupling is not a measurement oddity: an intervention that reduces internal shift by up to eight times in every tested condition simultaneously worsens output shift, revealing a genuine tension between internal and output-level fairness.

What carries the argument

The central object is the counterfactual identity probe: a pair of prompts identical in preference history and task instruction except for one protected-attribute sentence, such as "The user is a woman" versus "The user is a man." Each probe produces two ranked recommendation lists and two hidden-state sequences, and from these FairGap computes the output shift as one minus rank-biased overlap and the internal shift as a probe-weight-averaged cosine distance across four relative layer depths. The connection between the two axes is summarized by ROA, the Spearman correlation across users, and turned into a per-user diagnosis by adaptive two-dimensional Otsu thresholds that partition the joint shift plane into four quadrants. Activation steering on the top-5 gender-separable layers acts as the mechanical check: it confirms the internal axis is linearly decodable and shows that compressing it can destabilize outputs.

What would settle it

Recompute ROA and quadrant assignments across the full benchmark using cue-span mean and prompt-mean pooling instead of the last-prompt-token hidden state, holding everything else fixed, as the paper does for Llama-8B on MovieLens. If the low-ROA pattern and the non-negligible Q4 population disappear, change sign, or shrink to noise under these alternative anchors, then the claimed hidden-output decoupling is a property of the extraction rule, not of the models.

Watch

Extended reading notes

Core claim

The central claim is that output-level fairness evaluation of LLM recommenders is structurally incomplete: the ranked lists a model emits and the hidden states that produce them frequently disalign under a single protected-attribute cue. FairGap operationalizes this with Output Bias Score, defined as one minus rank-biased overlap between counterfactual top-10 lists; Internal Bias Score, defined as a separability-weighted average of cosine distances between hidden states at four relative layer depths; and Representation-Output Alignment, the Spearman correlation between the two across users. It then assigns each user pair to one of four quadrants of the joint shift space, isolating output-visible mismatch, joint sensitivity, joint stability, and hidden-internal mismatch, the last being invisible to output-only audits. The paper reports that ROA rarely exceeds 0.22 across the benchmark and that Q4 holds a non-negligible user population, and its steering experiments show that projection-mode activation steering reduces IBS by 3-8x while increasing OBS in every condition, which it reads as evidence of a fundamental tension rather than a measurement artifact.

Load-bearing premise

The load-bearing measurement assumption is that the hidden state at the last prompt token, sampled at four relative layer depths and compared by cosine distance, faithfully captures fairness-relevant internal processing; the paper's own ablation shows ROA and IBS change substantially under alternative extraction rules, so a wrong anchor would make the decoupling and Q4 populations artifacts of the extraction rule.

Editorial extensions

If this is right

  • An auditor who looks only at recommendation outputs can certify a model as fair while the model silently shifts its internal processing; benchmark results show Q4 populations persist across multiple model-domain pairs.
  • Utility metrics such as Match@10 do not certify fairness stability, since models with strong preference alignment can still show large internal displacement under attribute probes.
  • Debiasing interventions should not be assumed to improve both axes at once: steering that suppresses linear demographic directions can worsen observable divergence, so joint evaluation is needed to interpret intervention results.
  • New models can be compared against FairGap's standardized OBS, IBS, ROA, and quadrant profiles across the six model families and three domains.

Reading between the lines

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

  • If this paper is right, any output-only fairness certification for LLM recommenders is incomplete, and future audit standards may need to specify which internal representation anchor is being used, since the paper's own ablation shows ROA can flip sign across extraction rules.
  • The steering paradox suggests that removing a linearly decodable demographic direction is not sufficient for output fairness; a natural testable extension is steering on nonlinear or layer-targeted directions to see whether joint IBS/OBS improvement is possible.
  • Because IBS magnitudes are much smaller than OBS magnitudes, raw internal-shift numbers may be less interpretable across models than quadrant profiles; a practical extension is to report anchor-specific confidence intervals rather than point thresholds.
  • The same benchmark design could transfer to closed-weight models, but only where hidden-state access is available, so API-based audits would need a proxy internal signal or a stated limitation.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. FairGap introduces a benchmark for jointly evaluating LLM recommender fairness at the output level (OBS, based on rank-biased overlap of counterfactual recommendation lists) and at the internal-representation level (IBS, based on cosine distances between hidden states aggregated over four relative layer depths). The paper reports Representation-Output Alignment (ROA) as a Spearman correlation between these two axes, a four-quadrant taxonomy over the joint IBS-OBS space, and a steering intervention that is claimed to reduce IBS while worsening OBS. The central claims are that output-only fairness evaluation is structurally incomplete, that hidden-output decoupling is pervasive (with ROA rarely exceeding 0.22), and that a non-negligible population of users shows stable outputs despite substantial internal shifts. The paper also argues that activation steering demonstrates a fundamental tension between internal and output-level fairness.

Significance. If the central claims held, FairGap would be a useful contribution: the counterfactual minimal-pair construction is programmatically verified, the benchmark covers six model families across three domains and three attributes, and the paper includes several robustness checks (RBO persistence, percentile thresholding, prompt-family variants, age-cue variants). The release of benchmark data, probes, and evaluation code is a concrete strength. However, the load-bearing internal-representation measurement is not stable under alternative extraction anchors in the paper's own ablation, and the steering evidence is partly circular because the same representations define both the steering direction and the outcome metric. Because every downstream claim (low ROA, Q4 silent-bias population, steering paradox) is a comparison against this IBS measurement, the headline claims are not yet established as properties of the models. The framework is promising and the manuscript is repairable, but the abstract and Section 5 substantially oversell what the current evidence supports.

major comments (4)
  1. [Abstract and §5.1 (Table 3)] The claim that "ROA rarely exceeds 0.22" is contradicted by the paper's own Table 3. Ten of the 54 reported condition cells exceed 0.22, including Qwen-7B/MovieLens/Gender (0.315), Mistral-7B/SteamReviews/Age (0.379), Llama-8B/SteamReviews/Age (0.267), Dolphin-34B/SteamReviews/Gender (0.267), and four MovieLens/Race cells (0.235, 0.286, 0.289, 0.370). Even the cross-model average for MovieLens/Race is 0.224, itself above the stated threshold. This is not a minor wording issue: the abstract's central quantitative claim and the "pervasive decoupling" narrative are not supported by the reported distribution. The manuscript should either report the actual distribution of ROA values or revise the claim to match the data.
  2. [§A.3 (Table 6)] The extraction ablation in Appendix A.3 shows that IBS and ROA are highly sensitive to the choice of hidden-state anchor. For Llama-8B/MovieLens/Gender, ROA ranges from -0.244 (cue-span mean) to 0.181 (prompt-mean pooling), and IBS ranges from 4.59e-5 to 0.0398, a variation of nearly three orders of magnitude; for Race, ROA ranges from -0.104 to 0.289. Since IBS is the internal axis used to define Q4, to compute every ROA value, and to motivate the steering experiment, the headline decoupling phenomenon and the "silent bias" population are properties of the last-prompt-token extraction rule, not established model properties. The paper acknowledges this in A.3, where it concludes that the results support "a measurement-sensitive interpretation," but the Abstract and Section 5 present decoupling as a discovered property of LLM recommenders. This internal tension is load-bearing and must be resolved, either by demonstrating robustness across anchors or by explicitly reframing all conclusions as protocol-dependent.
  3. [§5.4 and §C.6] The steering experiment is partly circular. The steering direction is the normalized mean difference between female and male hidden states at each layer, the layer selection uses separability defined as the cosine distance between these class means, and the outcome IBS is a cosine-distance-based measure on the same hidden states. Projecting out the mean-difference vector from those representations will reduce the component of variance along that direction, so a drop in IBS is expected by construction and does not independently confirm that a demographic direction is "mechanistically present." The manuscript needs a control condition, for example steering with a random or permuted direction of the same norm, or evaluating IBS on held-out layers or held-out users not used to construct the vector, before the 3-8x reduction can be claimed as evidence of a genuine mechanistic effect.
  4. [Table 5 and §5.4] Table 5 presents six representative intervention conditions, but three rows are marked as estimated (‡) with methodology deferred to Appendix C.6, while the measured steering results in Table 12 cover only three conditions. The main text nevertheless states that "in every tested condition" OBS increases and describes this as "the strongest evidence" for FairGap's diagnostic utility. The abstract's "up to 8x" reduction is based on one measured row, but the generalization to all tested conditions relies on estimated rows. The manuscript should separate measured from estimated results in the main text and should restrict the "every tested condition" claim to the measured conditions, or move the estimated rows entirely to the appendix.
minor comments (5)
  1. [Section 6] The sentence "Although IBS is in absolute magnitude and diagnostically informative" is grammatically incomplete; it appears to be missing a predicate such as "is small in absolute magnitude."
  2. [Appendix C.6] The citation "[18,22]" for activation steering appears to point to CrowS-Pairs and a fairness-survey reference rather than the intended representation-steering works; it should likely be [34] and [44].
  3. [Table 3 caption] The caption states "Darker shading indicates more favourable values," but the table as rendered contains no shading; either implement the shading or remove the caption sentence.
  4. [Appendix A.3 vs B.1] The extraction ablation text uses the example "a Black user/a White user," while Section B.1 specifies the race cue as "a Black individual/a white individual"; the wording should be aligned to avoid confusion about which cue was actually tested.
  5. [Section 3.2] The mapping from relative depths L={1/4,2/4,3/4,4/4} to concrete layer indices is not specified for models where the layer count is not divisible by four; an off-by-one ambiguity could affect the layerwise results in Table 13.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: core decoupling metrics are not fitted to their conclusions; the steering IBS reduction is a manipulation check, and the paper's own anchor ablation is a validity caveat rather than a circular step.

full rationale

The main derivation is self-contained in the relevant sense. OBS (1 - RBO@K) is a standard rank-list distance, IBS is a probe-weighted cosine distance with weights fixed on a development split, and ROA is a Spearman correlation computed on a held-out evaluation split; none of these quantities is defined in terms of the paper's conclusions (low ROA, non-negligible Q4). The quadrant thresholds are adaptive, but the paper reports percentile-threshold robustness (Appendix C.4) showing Q4 remains populated, so the Q4 finding is not a threshold artifact. The steering experiment is the only candidate for circularity: the intervention projects out a mean-difference direction and selects layers by the same separability criterion used to weigh IBS. Because the projection removes the fitted component at the targeted layers, part of the observed IBS reduction is a manipulation check rather than an independent discovery. However, the steering vectors are computed on a training split and the IBS reduction is measured on a held-out evaluation split, and cosine distance after projection is not equationally guaranteed to decrease; the aggregate reduction is an empirical outcome. More importantly, the paper's central decoupling evidence does not rest on the steering result: the OBS increase under steering is independent and unforced, and the low-ROA/Q4 claims stand on the benchmark measurements. Appendix A.3 is a genuine limitation: the paper itself reports that ROA for Llama-8B/MovieLens/Gender ranges from -0.244 (cue-span mean) to +0.064 (last-prompt-token) to +0.181 (prompt-mean pooling) and endorses a 'measurement-sensitive interpretation.' That undermines the strength of the 'pervasive hidden-output decoupling' headline, but it is a construct-validity threat about which anchor to trust, not a case where a prediction is equivalent to its inputs by construction. There is no load-bearing self-citation, no imported uniqueness theorem, and no renamed known result; the counterfactual framing from prior fairness benchmarks is not used to prove the decoupling. Overall, the derivation is not circular; the appropriate concern is measurement robustness, which is a correctness issue, not a circularity issue.

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

The central metrics OBS, IBS, and ROA rest on several hand-chosen and fitted quantities. The most consequential is the representation anchor, whose alternative choices produce contradictory ROA signs in the paper's own Table 6; the fitted layer-separability weights and adaptive thresholds are secondary but still influence the headline numbers.

free parameters (5)
  • RBO persistence parameter p = 0.9
    Chosen by hand in Appendix A.1; sensitivity checks with 0.8 and 0.95 show magnitude shifts but the qualitative pattern is reported as stable.
  • Top-K list length K = 10
    Set to 10 for all experiments; affects both OBS and Match@10 and is a hand-chosen benchmark parameter.
  • Layerwise separability weights sep_ell = Varies per model, attribute, and domain (AUC-derived)
    Fitted on the development split via linear probe, then used to weight layerwise cosine distances in IBS; this fitted quantity directly shapes the central internal-shift metric.
  • Joint Otsu thresholds (t*_x, t*_y) = Per condition
    Estimated from the empirical joint IBS/OBS distribution for each model-dataset-attribute condition; quadrant assignments and Q4 prevalence claims depend on these thresholds.
  • Relative layer positions L = {1/4, 2/4, 3/4, 4/4}
    Hand-selected sampling of model depth; layerwise results in Appendix C.7 show non-monotonicity, so the aggregated IBS may depend on this choice.
assumptions (5)
  • domain assumption Counterfactual identity probes isolate the protected attribute: changing only the cue sentence, with all other prompt content byte-identical, measures the attribute's causal effect on outputs and hidden states.
    Section B.2 programmatically verifies the minimal-pair invariant, but assumes LLM behavior under greedy decoding is a stable function of prompt content and that no other uncontrolled variation exists.
  • ad hoc to paper Last-prompt-token hidden state is a meaningful representation of the model's internal processing for fairness measurement.
    Section A.3 selects this anchor as a 'principled middle ground' after showing cue-span and prompt-mean pooling produce ROA values that differ in sign and magnitude; the central decoupling claim is contingent on this choice.
  • domain assumption Cosine distance on hidden states at four relative depths captures fairness-relevant internal shift.
    Used throughout Section 3.2; it is one of many possible internal distance measures and is not justified against alternatives.
  • domain assumption Linear probe separability on the development split is a valid basis for weighting layer contributions to IBS.
    Appendix A.2 defines sep_ell as excess AUC over chance; assumes that layers where the attribute is linearly decodable are the layers that matter for fairness.
  • domain assumption The three source datasets (Goodreads, MovieLens, Steam) provide valid user preference profiles for recommendation, and synthetic demographic cues are an acceptable substitute for real demographic data.
    Section 4.1 constructs profiles from interaction histories without demographic ground truth; the counterfactual is about stated identity, not actual identity.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Fair on the Surface? Benchmarking Hidden-Output Fairness Gaps in LLM Recommenders." pith.science (2026). https://pith.science/paper/FTDLLJLC

@misc{pith2026260808284,
  author       = {Pith},
  title        = {Pith review of: Fair on the Surface? Benchmarking Hidden-Output Fairness Gaps in LLM Recommenders},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FTDLLJLC}},
  note         = {Machine review of arXiv:2608.08284}
}
read the original abstract

Fairness audits for LLM-based recommenders have largely focused on observable outputs, implicitly assuming that stable recommendations reflect stable internal processing. We challenge this assumption with FairGap, the first benchmark to jointly evaluate recommendation fairness at two levels: observable output shift (OBS) and hidden representation shift (IBS), measured through controlled counterfactual identity probes across gender, age, and race. Their relationship is summarized via Representation-Output Alignment (ROA), with quadrant diagnostics for identifying user-level hidden-output mismatch. Applied to six open-weight LLM families across three domains, FairGap reveals pervasive hidden-output decoupling: ROA rarely exceeds 0.22, and a non-negligible user population shows stable outputs despite substantial internal shifts, a mode that output-only audits cannot detect by design. Further, activation steering that reduces IBS by up to 8x simultaneously worsens OBS, demonstrating a fundamental tension between internal and output-level fairness that existing frameworks are unequipped to diagnose.

Figures

Figures reproduced from arXiv: 2608.08284 by the authors.

Figure 1
Figure 1. Motivation for FairGap. Two counterfactual recommendation prompts keep the same preference history and task, changing only the gender cue. Output-only fairness evaluation cannot distinguish between two qualitatively different cases: one in which recommendation outputs diverge despite similar internal representations, and another in which outputs remain similar enough to appear fair on the surface despite marked inte… view at source ↗
Figure 2
Figure 2. FairGap quadrant taxonomy. Adaptive thresh￾olds (t ∗ x , t∗ y ) partition the joint space of output shift (dout) and internal shift (din) into four diagnostic regions. FairGap organizes user-level counterfactual pairs into a four-quadrant taxonomy over the joint space of (dout, din) ( [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. FairGap benchmark construction pipeline. Fixed user profiles are converted into matched counterfactual prompt pairs that differ only in a protected-attribute cue. These pairs are evaluated with a shared LLM recommendation protocol to produce top-10 recommendation lists and internal representations, which are then used to compute utility, fairness, alignment, and quadrant assignments. 4 Benchmark Construction 4.1 Dom… view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: User-level internal–output decoupling under MovieLens gender counterfactuals [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Quadrant distributions under race counterfactuals. Each panel reports the proportion of user-level pairs assigned to one IBS–OBS quadrant for MovieLens and Goodreads across six LLMs. Q1 denotes output-visible mismatch, Q2 joint sensitivity, Q3 joint stability, and Q4 h…
Figure 6
Figure 6. Figure 6: Robustness of fairness metrics to prompt reformulation. OBS, IBS, and ROA under simple, structured, and optimized prompt families for MovieLens / Gender / Llama-3.1-8B. Prompt wording affects magnitude but does not remove the hidden–output fairness gap. B.4 Compute res…
Figure 7
Figure 7. Figure 7: Robustness of quadrant-level patterns to prompt reformulation. Quadrant distributions under three prompt families for the same MovieLens / Gender / Llama-3.1-8B condition. Relative frequencies of Q1–Q4 vary across formulations, but mismatch regions remain populated, su…
Figure 8
Figure 8. Figure 8: Locations of representative case studies in the IBS–OBS space. Each point is a user-level counterfactual pair for Llama-8B on MovieLens. Dashed lines denote Joint Otsu thresholds. Case 1: output-visible mismatch; Case 2: hidden-internal mismatch. C.2 Utility–fairness a…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

44 extracted references · 25 canonical work pages

  1. [1]

    O’Brien, Carrie J

    Joon Sung Park, Joseph C. O’Brien, Carrie J. Cai, Meredith Ringel Morris, Percy Liang, and Michael S. Bernstein. Generative agents: Interactive simulacra of human behavior. In Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology. ACM, 2023. doi: 10.1145/3586183.3606763

    arXiv 2023

  2. [2]

    The rise and potential of large language model based agents: A survey.Science China Information Sciences, 68, 2025

    Zhiheng Xi, Wenxiang Chen, Xin Guo, Wei He, Yiwen Ding, Boyang Hong, Ming Zhang, Junzhe Wang, Senjie Jin, Enyu Zhou, Rui Zheng, Xiaoran Fan, Xiao Wang, Limao Xiong, Yuhao Zhou, Weiran Wang, Changhao Jiang, Yicheng Zou, Xiangyang Liu, Zhangyue Yin, Shihan Dou, Rongxiang Weng, Wensen Cheng, Qi Zhang, Wenjuan Qin, Yongyan Zheng, Xipeng Qiu, Xuanjing Huang, a...

    work page 2025

  3. [3]

    WebGPT: Browser-assisted question-answering with human feedback.arXiv preprint arXiv:2112.09332, 2021

    Reiichiro Nakano, Jacob Hilton, Suchir Balaji, Jeff Wu, Long Ouyang, Christina Kim, Christo- pher Hesse, Shantanu Jain, Vineet Kosaraju, William Saunders, Xu Jiang, Karl Cobbe, Tyna Eloundou, Gretchen Krueger, Kevin Button, Matthew Knight, Benjamin Chess, and John Schul- man. WebGPT: Browser-assisted question-answering with human feedback.arXiv preprint a...

    arXiv 2021

  4. [4]

    Large language models for information retrieval: A survey.ACM Transactions on Information Systems, 2025

    Yutao Zhu, Huaying Yuan, Shuting Wang, Jiongnan Liu, Wenhan Liu, Chenlong Deng, Haonan Chen, Zheng Liu, Zhicheng Dou, and Ji-Rong Wen. Large language models for information retrieval: A survey.ACM Transactions on Information Systems, 2025. doi: 10.1145/3748304

    doi:10.1145/3748304 2025

  5. [5]

    Recommender systems in the era of large language models.IEEE Transactions on Knowledge and Data Engineering, 2024

    Zihuai Zhao, Wenqi Fan, Jiatong Li, Yunqing Liu, Xiaowei Mei, Yiqi Wang, Zhen Wen, Fei Wang, Xiangyu Zhao, Jiliang Tang, and Qing Li. Recommender systems in the era of large language models.IEEE Transactions on Knowledge and Data Engineering, 2024. doi: 10.1109/TKDE.2024.3392335

    arXiv 2024

  6. [6]

    Large language models for generative recommendation: A survey and visionary discussions.arXiv preprint arXiv:2309.01157, 2023

    Lei Li, Yongfeng Zhang, Dugang Liu, and Li Chen. Large language models for generative recommendation: A survey and visionary discussions.arXiv preprint arXiv:2309.01157, 2023

    arXiv 2023

  7. [7]

    Zhang, K

    J. Zhang, K. Bao, Y . Zhang, W. Wang, F. Feng, and X. He. Is ChatGPT fair for recommendation? Evaluating fairness in large language model recommendation. InProceedings of the 17th ACM Conference on Recommender Systems, pages 993–999. ACM, 2023

    work page 2023

  8. [8]

    W. Hua, Y . Ge, S. Xu, J. Ji, and Y . Zhang. UP5: Unbiased foundation model for fairness-aware recommendation. InProceedings of the 18th Conference of the European Chapter of the Association for Computational Linguistics, pages 457–473. Association for Computational Linguistics, 2024

    work page 2024

Show all 44 references
  1. [9]

    Y . Li, M. Du, R. Song, X. Wang, and Y . Wang. A survey on fairness in large language models. arXiv preprint arXiv:2308.10149, 2023

    2023 arXiv

  2. [10]

    Z. Chu, Z. Wang, and W. Zhang. Fairness in large language models: A taxonomic survey. SIGKDD Explorations Newsletter, 26(1):34–48, 2024

    2024

  3. [11]

    T.V . Doan, Z. Chu, Z. Wang, and W. Zhang. Fairness definitions in language models explained. arXiv preprint arXiv:2407.18454, 2024

    2024

  4. [12]

    Griffiths

    Xuechunzi Bai, Angelina Wang, Ilia Sucholutsky, and Thomas L. Griffiths. Explicitly unbiased large language models still form biased associations.Proceedings of the National Academy of Sciences, 122(8):e2416228122, 2025. doi: 10.1073/pnas.2416228122

    2025 doi

  5. [13]

    Aligned but blind: Align- ment increases implicit bias by reducing awareness of race

    Lihao Sun, Chengzhi Mao, Valentin Hofmann, and Xuechunzi Bai. Aligned but blind: Align- ment increases implicit bias by reducing awareness of race. InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics, pages 22167–22184. Association for Com...

    2025

  6. [14]

    Cassese, G

    M. Cassese, G. Puccetti, and A. Esuli. Prompt-based bias control in large language models: A mechanistic analysis. InHHAI-WS 2025: Workshops at the Fourth International Conference on Hybrid Human-Artificial Intelligence, 2025

    2025

  7. [15]

    A. Goel, A. Ritter, and I. Gurevych. Auditing language model unlearning via information decomposition.arXiv preprint arXiv:2601.15111, 2026

    2026

  8. [16]

    Hernandez, B.Z

    E. Hernandez, B.Z. Li, and J. Andreas. Inspecting and editing knowledge representations in language models. InProceedings of the Conference on Language Modeling (COLM 2024), 2024

    2024

  9. [17]

    Gender bias in coreference resolution: Evaluation and debiasing methods

    Jieyu Zhao, Tianlu Wang, Mark Yatskar, Vicente Ordonez, and Kai-Wei Chang. Gender bias in coreference resolution: Evaluation and debiasing methods. InProceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Langua...

    2018

  10. [18]

    Nikita Nangia, Clara Vania, Rasika Bhalerao, and Samuel R. Bowman. CrowS-Pairs: A challenge dataset for measuring social biases in masked language models. InProceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, pages 1953–1967. Association for...

    2020

  11. [19]

    Alicia Parrish, Angelica Chen, Nikita Nangia, Vishakh Padmakumar, Jason Phang, Jana Thomp- son, Phu Mon Htut, and Samuel R. Bowman. BBQ: A hand-built bias benchmark for question answering. InFindings of the Association for Computational Linguistics: ACL 2022, pages 2086–2105. ...

    2022

  12. [20]

    Samuel Gehman, Suchin Gururangan, Maarten Sap, Yejin Choi, and Noah A. Smith. Real- ToxicityPrompts: Evaluating neural toxic degeneration in language models. InFindings of the Association for Computational Linguistics: EMNLP 2020, pages 3356–3369. Association for Computational...

    2020

  13. [21]

    Greenwood, S

    S. Greenwood, S. Chiniah, and N. Garg. User-item fairness tradeoffs in recommendations. In Advances in Neural Information Processing Systems, volume 38, 2024

    2024

  14. [22]

    Y . Zhao, Y . Wang, Y . Liu, X. Cheng, C.C. Aggarwal, and T. Derr. Fairness and diversity in recommender systems: A survey.ACM Transactions on Intelligent Systems and Technology, 2024

    2024

  15. [23]

    Rampisela, M

    T.V . Rampisela, M. Maistro, T. Ruotsalo, F. Scholer, and C. Lioma. Stairway to fairness: Connecting group and individual fairness. InProceedings of the 19th ACM Conference on Recommender Systems, pages 1–7. ACM, 2025

    2025

  16. [24]

    Rampisela, M

    T.V . Rampisela, M. Maistro, T. Ruotsalo, and C. Lioma. Measuring individual user fairness with user similarity and effectiveness disparity.arXiv preprint arXiv:2602.02516, 2026

    2026

  17. [25]

    A. Aird, B. Armstrong, N. Mattei, and R. Burke. Envy-free but still unfair: Envy-freeness up to one item (EF-1) in personalized recommendation. InProceedings of the FAccTRec@RecSys’25 Workshop, pages 1–3. ACM, 2025

    2025

  18. [26]

    Is ChatGPT fair for recommendation? Evaluating fairness in large language model recommendation.arXiv preprint arXiv:2305.07609, 2023

    Jizhi Zhang, Keqin Bao, Yang Zhang, Wenjie Wang, Fuli Feng, and Xiangnan He. Is ChatGPT fair for recommendation? Evaluating fairness in large language model recommendation.arXiv preprint arXiv:2305.07609, 2023

    2023

  19. [27]

    Ethical and social risks of harm from language models.arXiv preprint arXiv:2112.04359, 2021

    Laura Weidinger, John Mellor, Maribeth Rauh, Conor Griffin, Jonathan Uesato, Po-Sen Huang, Myra Cheng, Mia Glaese, Borja Balle, Atoosa Kasirzadeh, et al. Ethical and social risks of harm from language models.arXiv preprint arXiv:2112.04359, 2021

    2021 arXiv

  20. [28]

    Gallegos, Ryan A

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

    2024

  21. [29]

    BOLD: Dataset and metrics for measuring biases in open-ended language generation

    Jwala Dhamala, Tony Sun, Varun Kumar, Satyapriya Krishna, Yada Pruksachatkun, Kai-Wei Chang, and Rahul Gupta. BOLD: Dataset and metrics for measuring biases in open-ended language generation. InProceedings of FAccT, pages 862–872, 2021. 11

    2021

  22. [30]

    T.V . Doan, Z. Wang, N.N.M. Hoang, and W. Zhang. Fairness in large language models in three hours. InProceedings of the 33rd ACM International Conference on Information and Knowledge Management, pages 5514–5517. ACM, 2024

    2024

  23. [31]

    Y .O. Wang, N. Sivakumar, F.A. Khan, R.M. Susa, A. Golinski, N. Mackraz, B.-J. Theobald, L. Zappella, and N. Apostoloff. Is your model fairly certain? Uncertainty-aware fairness evaluation for LLMs. InProceedings of the 42nd International Conference on Machine Learning, volume...

    2025

  24. [32]

    Z. Fan, R. Chen, T. Hu, and Z. Liu. FairMT-Bench: Benchmarking fairness for multi-turn dialogue in conversational LLMs. InInternational Conference on Learning Representations, 2025

    2025

  25. [33]

    Iskander, K

    S. Iskander, K. Radinsky, and Y . Belinkov. Shielded representations: Protecting sensitive attributes through iterative gradient-based projection. InFindings of the Association for Compu- tational Linguistics: ACL 2023, pages 5961–5977. Association for Computational Linguistics, 2023

    2023

  26. [34]

    A. Zou, L. Phan, S. Chen, J. Campbell, P. Guo, R. Ren, A. Pan, X. Yin, M. Mazeika, A.-K. Dombrowski, S. Goel, N. Li, M.J. Byun, Z. Wang, A. Mallen, S. Basart, S. Koyejo, D. Song, M. Fredrikson, Z. Kolter, and D. Hendrycks. Representation engineering: A top-down approach to AI ...

    arXiv

  27. [35]

    Golgoon, K

    A. Golgoon, K. Filom, and A.R. Kannan. Mechanistic interpretability of large language models with applications to the financial services industry. InProceedings of the 5th ACM International Conference on AI in Finance, pages 660–668. ACM, 2024

    2024

  28. [36]

    Cohen-Wang, H

    B. Cohen-Wang, H. Shah, K. Georgiev, and A. M ˛ adry. CONTEXTCITE: Attributing model generation to context.arXiv preprint arXiv:2409.00729, 2024

    2024 arXiv

  29. [37]

    Y . Wang, R. Geng, Y . Chen, and J. Jia. AttnTrace: Attention-based context traceback for long-context LLMs.arXiv preprint arXiv:2508.03793, 2025

    2025 arXiv

  30. [38]

    Webber, A

    W. Webber, A. Moffat, and J. Zobel. A similarity measure for indefinite rankings.ACM Transactions on Information Systems, 28(4):1–38, 2010

    2010

  31. [39]

    A threshold selection method from gray-level histograms.IEEE Transactions on Systems, Man, and Cybernetics, 9(1):62–66, 1979

    Nobuyuki Otsu. A threshold selection method from gray-level histograms.IEEE Transactions on Systems, Man, and Cybernetics, 9(1):62–66, 1979. doi: 10.1109/TSMC.1979.4310076

    1979

  32. [40]

    Item recommendation on monotonic behavior chains

    Mengting Wan and Julian McAuley. Item recommendation on monotonic behavior chains. In Proceedings of the 12th ACM Conference on Recommender Systems, pages 86–94. Association for Computing Machinery, 2018. doi: 10.1145/3240323.3240369

    2018

  33. [41]

    Maxwell Harper and Joseph A

    F. Maxwell Harper and Joseph A. Konstan. The movielens datasets: History and context.ACM Transactions on Interactive Intelligent Systems, 5(4):19:1–19:19, 2015. doi: 10.1145/2827872

    2015 doi

  34. [42]

    Game recommendations on steam

    Anton Kozyriev. Game recommendations on steam. Kaggle dataset, 2021. Available at: https: //www.kaggle.com/datasets/antonkozyriev/game-recommendations-on-steam

    2021

  35. [43]

    System prompt optimization with meta- learning.arXiv preprint arXiv:2505.09666, 2025

    Yumin Choi, Jinheon Baek, and Sung Ju Hwang. System prompt optimization with meta- learning.arXiv preprint arXiv:2505.09666, 2025

    2025

  36. [44]

    The user is a woman

    A. Turner, L. Thierauf, N. Goldowsky-Dill, C. McDougall, J. Rao, A. Mallen, N. Nanda, and T. Lieberum. Activation addition: Steering language models without optimization.arXiv preprint arXiv:2308.10248, 2023. A Metric definitions and implementation details A.1 Rank-biased over...

    2023 arXiv

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.