Pith. sign in

REVIEW 4 major objections 4 minor 1 cited by

FACTER: Fairness-Aware Conformal Thresholding and Prompt Engineering for Enabling Fair LLM-Based Recommender Systems

T0 review · 4 major / 4 minor · reviewed 2026-08-09 · deepseek-v4-flash

Pith's one-line read Conformal thresholds on semantic variance, paired with violation-triggered prompt repair, cut fairness violations in LLM recommenders by up to 95.5 percent while preserving accuracy.

desk verdict The paper's central adaptive conformal threshold is a mathematical no-op, so the headline mechanism does not exist; the prompt-repair idea underneath is real but the submission fails on its own terms. read the letter →

arxiv 2502.02966 v1 pith:VSNSVLMX submitted 2025-02-05 cs.IR cs.AIcs.CYcs.LG

classification cs.IRcs.AIcs.CYcs.LG
keywords fairness-awarerecommendationconformalpredictionpromptengineeringLLMbiasmitigationsemanticvarianceblack-boxcounterfactualfairnessrecommendersystems
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

FACTER claims that fairness in black-box LLM recommenders can be enforced without retraining, by using conformal prediction to set a threshold on semantic variance and by rewriting the prompt whenever the threshold is crossed. On MovieLens-1M and Amazon Movies & TV, the authors report that the iterative procedure cuts fairness violations by up to 95.5 percent while keeping recommendation accuracy competitive with, and sometimes above, fairness-aware baselines such as UP5. The central move is to treat a large embedding-space divergence between responses to minimally changed demographics as a fairness violation, then inject the offending context-attribute-output triple as an explicit “avoid” example into the system prompt. If the claim holds, post hoc prompt repair is a viable replacement for retraining in LLM-based recommendation fairness.

What carries the argument

The load-bearing object is the fairness-aware non-conformity score $S_i = d_i + \lambda \Delta_i$, where $d_i$ is the cosine distance between the LLM's output embedding and the reference item embedding, and $\Delta_i$ is the maximum embedding distance to calibration peers whose contexts are similar ($W_{ij} > \tau_\rho$) but whose protected attributes differ. A conformal quantile over calibration scores sets the initial violation threshold, and whenever a new query's score exceeds it the framework stores the tuple, appends explicit negative patterns to the system prompt, and shrinks the threshold geometrically. The named mechanism is FACTER's violation-triggered prompt repair loop, which is what lets a black-box model improve fairness without weight updates.

What would settle it

Take a deployment log and replace the oracle reference embedding $e^y_{\mathrm{new}}$ with the embedding of the item each user actually consumed next, then replay the online phase. If the violation counts no longer fall monotonically across iterations, or the gap between FACTER and UP5 narrows to near zero, the claimed 95.5 percent reduction depends on access to the ideal item rather than on the prompt-repair mechanism itself.

Watch

Extended reading notes

Core claim

The paper's core claim is that a closed loop of conformal thresholding and prompt repair reduces demographic bias in LLM recommenders: a calibration set yields an initial conformal quantile $Q^{(0)}_\alpha$ over fairness-aware non-conformity scores $S_i = d_i + \lambda \Delta_i$, where $d_i$ is the embedding distance from the recommended item to the reference item and $\Delta_i$ penalizes divergence from similar-context, different-attribute peer outputs. Online, each new output is scored the same way and compared with the current threshold; scores above it trigger a FIFO memory of violations, injection of explicit “avoid (context, attribute) → output” patterns into the system prompt, and an exponential threshold decay $Q^{(t+1)}_\alpha = \gamma Q^{(t)}_\alpha + (1-\gamma)\min(Q^{(t)}_\alpha, S_{\mathrm{new}})$. The authors report that on MovieLens-1M this reduces fairness violations from 112 for zero-shot ranking and 28 for UP5 to 5 at the third iteration, with NDCG@10 of 0.445 versus 0.427 for UP5; on the sparser Amazon Movies & TV data, violations fall from 198 to 18 with CFR dropping from 0.814 to 0.634. The paper further claims that explicit pattern enumeration in the prompt update outperforms generic warnings or single negative examples.

Load-bearing premise

The online fairness check assumes it knows the embedding of the ideal or reference output for each new user query, yet a deployed recommender does not know in advance which item the user ideally wanted; without that oracle, $S_{\mathrm{new}}$ cannot be computed and the violation-detection loop cannot start.

Editorial extensions

If this is right

  • Fairness violations can be reduced by an order of magnitude relative to zero-shot LLM ranking and by roughly 80 percent relative to UP5, according to the MovieLens experiment, with NDCG@10 slightly above UP5.
  • The same closed loop transfers to sparser data: on Amazon Movies & TV violations drop from 198 to 18 and CFR falls from 0.814 to 0.634, indicating the method does not rely on dense interaction data.
  • The procedure works across LLMs of different sizes (LLaMA3-8B, LLaMA2-7B, Mistral-7B), with all three converging near zero violations by iteration 3, so the mechanism is not tied to one model's prompt behavior.
  • Explicit enumeration of repeated bias patterns in the prompt is the operative repair strategy; generic warnings without concrete examples leave most violations in place.
  • The conformal construction supplies a bound on the Type I error rate of false violation alarms, so the threshold is a statistical object rather than an arbitrary hand-set value.

Reading between the lines

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

  • The online scoring step needs the embedding of the ideal or reference output for each new query (Eq. 9, with the phrase “if available”); in a real deployed recommender that ground truth is unknown. A testable extension is to compute the online score against the item the user actually consumed next in logged data, and to check whether the 95.5 percent reduction survives that substitution.
  • Semantic variance as a proxy for bias suggests a portable detector: any black-box text generator could be wrapped with the same threshold-and-repair loop, not just recommenders, as long as an embedder and a calibration set with paired demographic flips exist.
  • The threshold decay and prompt injection act through different channels—one compresses what counts as acceptable divergence, the other tells the model what to avoid—so an ablation that removes either component could reveal whether their joint effect is additive or whether one dominates the violation reduction.
  • If the embedder itself encodes demographic stereotypes, the fairness signal is corrupted before thresholding; swapping in several embedders and comparing violation trajectories would test how much of the reported reduction is inherited from the embedding model.
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 / 4 minor

Summary. The paper proposes FACTER, a two-phase framework for making black-box LLM-based recommenders fairer without retraining. Offline, it computes semantic non-conformity scores from a calibration set and sets a conformal quantile threshold. Online, it scores each recommendation against the current threshold and, on violation, stores the sample, injects an 'avoid' example into the system prompt, and purportedly tightens the threshold via an exponential decay update. Experiments on MovieLens-1M and Amazon Movies & TV report up to a 95.5% reduction in fairness violations with little accuracy loss, supported by theoretical bounds in the appendix.

Significance. If the mechanism worked as described, FACTER would be practically valuable: it is retraining-free, black-box compatible, and combines statistical calibration with prompt repair for LLM recommenders. The paper also provides ablations over prompt strategies and hyperparameters. However, the central adaptive threshold update is a mathematical no-op, the online fairness score requires an unobservable ground-truth item, and the evaluation uses the same embedding distances that the intervention directly minimizes. These issues substantially undermine the paper's central claims.

major comments (4)
  1. [Section 3.3, Eq. (11); Algorithm 1 line 12] The threshold update in Eq. (11) never changes Q on a violation. A violation is defined by S_new > Q_t; in that case min(Q_t, S_new) = Q_t, so Q_{t+1} = gamma Q_t + (1 - gamma) Q_t = Q_t. Thus the 'adaptive semantic variance threshold' is inert, and the text immediately after Eq. (11) claiming 'the threshold shrinks until we again achieve approximate alpha-coverage' is not implemented by the stated equation. Appendix Theorem .2 uses the different update Q_{t+1} = gamma Q_t + (1 - gamma) S_t when S_t > Q_t, which would increase the threshold on violations, again contradicting the paper's description. Because the abstract and Section 1 credit adaptive conformal thresholding as a core contribution, this is a load-bearing inconsistency.
  2. [Section 3.3, Stage 2, Eq. (9)] The real-time fairness score S_new = d_new + lambda Delta_new requires d_new = 1 - cos(Emb(yhat_new), e^{y_new}), where e^{y_new} is the embedding of the 'ideal or reference output' for the new query. In a deployed recommender, the ideal item for a new user request is not known, so S_new cannot be computed and violation detection cannot run online. The parenthetical 'if available' is not a fallback, and the experiments in Tables 1-3 necessarily use ground-truth references, meaning the reported online violation reduction is not reproducible in a deployment setting.
  3. [Section 3.2, Lemma 3.1; Section 3.3] The conformal coverage guarantee in Lemma 3.1 assumes that calibration points and future data are exchangeable. In FACTER's online phase, the prompt I(t) is modified after every detected violation, and these modifications depend on the observed sequence of violations. Future scores are therefore not exchangeable with the offline calibration set, so P(S_new <= Q_alpha^(0)) >= 1 - alpha does not hold under the online procedure. The paper provides no recalibration step or martingale-style argument to restore the guarantee.
  4. [Section 2.2, Eq. (5); Section 4.1, Eqs. (12)-(14)] The evaluation is substantially circular. Fairness violations are detected using embedding distances in the scorer of Eq. (5), and the prompt-repair mechanism is designed to reduce exactly those distances. The reported fairness metrics SNSR and CFR are also embedding- or counterfactual-distance based. Thus the observed reduction in violations and the improvement in SNSR/CFR are partly by construction, since the detector, the intervention target, and the evaluation measure all live in the same embedding space. Independent behavioral or human-validated fairness measures are needed to support the claim that FACTER improves fairness rather than merely reducing the specific distance used in both detection and evaluation.
minor comments (4)
  1. [Table 8 caption] The label 'FACTOR (no calibration)' appears to be a typo for 'FACTER'.
  2. [Section 4.3, Eq. (18)] The symbol delta is overloaded: it is used for the fairness tolerance in Eq. (1), for the confidence parameter in Eq. (18), and for embedding perturbation in Appendix Theorem .1. Please rename at least two of these.
  3. [Section 4.1, Eq. (15)] Equation (15) defines the violation threshold as Quantile(1-alpha; {s_i}) + C/sqrt(n), which is not reconciled with the offline conformal quantile in Eq. (6) or with the threshold update in Eq. (11). The finite-sample correction C is never defined or analyzed.
  4. [Section 4.2, Figures 3-4] The horizontal axis 'Iteration' is not precisely defined. It should state whether each iteration is a pass over the test set, a batch of online queries, or a calibration round, and how the violation count at each iteration is accumulated.

Circularity Check

2 steps flagged · score 6.0 of 10

The paper's central fairness proxy is definitional, and its claimed adaptive threshold update is a no-op by construction.

  1. self definitional [Section 2.2 Eq. (1); Section 3.2 Eq. (5); Abstract]
    "An LLM-based recommender ˆY : X × A → Y satisfies a minimal-attribute-change fairness property if altering only the sensitive attribute a → a′ (while holding x fixed) does not yield large discrepancies in the resulting outputs: ‖ ˆY (x, a) − ˆY (x, a′)‖ ≤ δ. (1) ... S_i = 1 − cos(Emb(ˆy_i), e^y_i) + λ max_{j: W_ij>τρ} ‖Emb(ˆy_i) − Emb(ˆy_j)‖_2 (5) ... Empirical results ... revealing semantic variance as a potent proxy of bias."

    The paper defines fairness (Eq. 1) as a small embedding distance under minimal attribute change, and the non-conformity score that flags violations (Eq. 5) is built from exactly those embedding distances. A violation is, by construction, a large semantic-variance score. Therefore the abstract's conclusion that semantic variance is a potent proxy of bias is not an independent empirical discovery; it is the same quantity used to define bias and to compute the reported violation counts. The measured reduction in violations is thus partly guaranteed by the definitional loop between the detector and the target metric.

  2. other [Section 3.3 Eq. (11); Algorithm 1 line 12; Appendix Theorem .2]
    "Q(t+1)_α = γ Q(t)_α + (1 − γ) min(Q(t)_α, S_new), where γ ∈ (0, 1). Smaller γ makes the threshold adapt more aggressively (i.e., decreasing it further whenever a violation is found). This dynamic ensures that if violations consistently appear, the threshold shrinks until we again achieve approximate α-coverage."

    The update is executed only when a violation is detected, i.e., when S_new > Q(t)_α. In that regime min(Q(t)_α, S_new) = Q(t)_α, so Q(t+1)_α = γ Q(t)_α + (1 − γ) Q(t)_α = Q(t)_α. When no violation occurs, Q is unchanged. Hence Q(t) is invariant for every possible input, and the advertised adaptive threshold reduces by construction to a constant threshold. The appendix's Theorem .2 uses a different update, Q(t+1)_α = γ Q(t)_α + (1 − γ) S_t when S_t > Q(t)_α, which would increase rather than shrink the threshold, so neither equation implements the claimed tightening mechanism.

full rationale

FACTER is not a case of self-citation circularity: the authors cite no prior work of their own, and the conformal machinery is credited to standard external references. The comparisons against UP5 and Zero-Shot on SNSR and CFR provide some independent content, because those baselines are measured under the same scoring protocol. However, the central fairness construct is self-referential: semantic variance is both the definition of bias (Eq. 1) and the ingredient in the violation score (Eq. 5), so reporting that semantic variance predicts bias and that FACTER reduces fairness violations is partially a restatement of the paper's own operationalization. In addition, the key adaptive-threshold contribution (Eq. 11 and Algorithm 1) is a mathematical no-op: on a violation, the min term equals the current threshold, so the threshold never changes. This is more than a correctness concern; it removes the only quantitative mechanism claimed to produce tightening, leaving the prompt-repair loop as the real intervention. The online scoring also requires the oracle embedding e^y_new of the ideal item, which is not available in deployment, though that is an operational limitation rather than a circularity. Overall, the definitional proxy claim and the inert threshold update justify a score of 6: some central 'predictions' reduce by construction, while the prompt-engineering component remains an empirically meaningful intervention.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

The framework's guarantees rest on exchangeability and oracle reference outputs, and the hyperparameters are tuned on calibration data. The score mixes accuracy and disparity, so the fairness-violation concept is partly defined by the same embedding geometry used for mitigation. The threshold update is internally inconsistent.

free parameters (7)
  • lambda (fairness penalty weight) = 0.7
    Balances accuracy and fairness in the score (Eq. 5); grid-selected (Table 5) and affects the central violation count and NDCG.
  • gamma (threshold decay factor) = 0.95
    Grid-selected; used in Eq. (11), but the update is a no-op under violations, so the empirical role is unexplained.
  • tau_rho (cross-group context similarity threshold) = 0.9
    Defines the local neighborhood in Eq. (4); grid-selected (Table 6).
  • tau_x (context-space radius) = not reported
    Appears in Eq. (4) as a condition for pairwise similarity; no value or tuning described.
  • M (FIFO buffer size) = 50
    Determines how many violation examples are kept for prompt injection; set by token-budget constraints.
  • C (finite-sample correction) = not reported
    Appears only in experimental threshold Eq. (15); unspecified and inconsistent with Eq. (6).
  • alpha (coverage level) = 0.2 (inferred)
    Implied by Table 4's Type I Error bound of 0.201; no explicit discussion of how alpha is chosen.
assumptions (5)
  • domain assumption Calibration and test samples are exchangeable (Lemma 3.1).
    Required for the conformal coverage guarantee in Eq. (7), but online prompt/threshold updates break exchangeability.
  • domain assumption An oracle or reference item y_new is available for every new query so the accuracy term d_new can be evaluated.
    Stage 2 (Eq. 9) requires e^y_new = Emb(y_new); the paper says 'if available' with no fallback, so the online monitor cannot run in deployment without ground truth.
  • domain assumption The embedding function Emb is relatively bias-free.
    Section 3.4 lists this as limitation 2; if embeddings carry bias, the semantic-distance proxy and calibration are compromised.
  • domain assumption Semantic embedding distance is a valid measure of unfairness (minimal-attribute-change fairness, Eq. 1).
    The fairness definition equates fairness with small embedding distances across protected attributes; this is a substantive modeling choice, not a mathematically forced one.
  • standard math Standard conformal quantile formula (Eq. 6) is valid.
    This is the textbook finite-sample quantile; accepted as a mathematical tool.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FACTER: Fairness-Aware Conformal Thresholding and Prompt Engineering for Enabling Fair LLM-Based Recommender Systems." pith.science (2026). https://pith.science/paper/VSNSVLMX

@misc{pith2026250202966,
  author       = {Pith},
  title        = {Pith review of: FACTER: Fairness-Aware Conformal Thresholding and Prompt Engineering for Enabling Fair LLM-Based Recommender Systems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VSNSVLMX}},
  note         = {Machine review of arXiv:2502.02966}
}
read the original abstract

We propose FACTER, a fairness-aware framework for LLM-based recommendation systems that integrates conformal prediction with dynamic prompt engineering. By introducing an adaptive semantic variance threshold and a violation-triggered mechanism, FACTER automatically tightens fairness constraints whenever biased patterns emerge. We further develop an adversarial prompt generator that leverages historical violations to reduce repeated demographic biases without retraining the LLM. Empirical results on MovieLens and Amazon show that FACTER substantially reduces fairness violations (up to 95.5%) while maintaining strong recommendation accuracy, revealing semantic variance as a potent proxy of bias.

Figures

Figures reproduced from arXiv: 2502.02966 by the authors.

Figure 1
Figure 1. FACTER’s Iterative Prompt Engineering in Practice. (1) A user requests movie rec￾ommendations. (2) The LLM response uses demographic information (‘30-year-old woman”) to suggest a stereotypical romance. (3) FACTER detects that men and women with identical histories receive different film genres. (4) FACTER inserts a new “avoid” example into the system prompt, indicating that having bias on gender is unacceptable (un… view at source ↗
Figure 2
Figure 2. FACTER Framework Workflow. The system operates in two coordinated phases: (Left) Offline calibration computes fairness-aware thresholds using historical data (Stages A–C): (A) Data prepro￾cessing and calibration, (B) Fairness scoring, and (C) Calculation of initial quantile thresholds. (Right) Online deployment with continuous monitoring (Stages 1–3): (1) New queries generate (2) LLM recommen￾dations that undergo (3… view at source ↗
Figure 3
Figure 3. Fairness violation reduction trajectory vs. [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Violation reduction across LLMs. All mod [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: Fairness-accuracy tradeoff comparison. FACTER achieves strong fairness improvement while preserving recommendation quality. 4.3 Theoretical Validation Beyond empirical performance, we provide theoret￾ical guarantees for our conformal calibration frame￾work. Our derivat…
Figure 6
Figure 6. Figure 6: Ablation on γ: final violation counts on Amazon over 5 calibration rounds, λ = 0.7. All converge by iteration 5, but larger γ lowers the threshold more gradually. .2.3 Neighborhood Similarity τρ Finally, we vary τρ ∈ {0.80, 0.85, 0.90, 0.95} in constructing local fairn…
Figure 7
Figure 7. Figure 7: Convergence of fairness violations over 5 calibration roun [PITH_FULL_IMAGE:figures/full_fig_p019_7.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

  1. MARCO: Hardware-Aware Neural Architecture Search for Edge Devices with Multi-Agent Reinforcement Learning and Conformal Prediction Filtering

    cs.LG 2025-06 conditional novelty 5.0 of 10

    A two-agent RL search with a conformal prediction filter designs mixed-precision networks for microcontrollers, cutting search time by 3-4x versus once-for-all with near-equal accuracy.

Reference graph

Works this paper leans on

35 extracted references · 19 canonical work pages · cited by 1 Pith paper

  1. [1]

    N., Bates, S., et al

    Angelopoulos, A. N., Bates, S., et al. Conformal prediction: A gentle introduction. Foundations and Trends in Machine Learning , 16 0 (4): 0 494--591, 2023

  2. [2]

    Training a helpful and harmless assistant with reinforcement learning from human feedback

    Bai, Y., Jones, A., Ndousse, K., Askell, A., Chen, A., DasSarma, N., Drain, D., Fort, S., Ganguli, D., Henighan, T., et al. Training a helpful and harmless assistant with reinforcement learning from human feedback. arXiv preprint arXiv:2204.05862, 2022

  3. [3]

    On accumulation points of volumes of stable surfaces with one cyclic quotient singularity

    Bary, R., Gehrmann, S., Nadeem, M., and Tabbal, A. We need to consider more than just race and gender: Bias in language generation extended to intersections. arXiv preprint arXiv:2107.01317, 2021

  4. [4]

    M., Gebru, T., McMillan-Major, A., and Shmitchell, S

    Bender, E. M., Gebru, T., McMillan-Major, A., and Shmitchell, S. On the dangers of stochastic parrots: Can language models be too big? In Proceedings of the 2021 ACM conference on fairness, accountability, and transparency, pp.\ 610--623, 2021

  5. [5]

    Language (technology) is power: A critical survey of "bias" in nlp

    Blodgett, S., Barocas, S., Daumé III, H., and Wallach, H. Language (technology) is power: A critical survey of "bias" in nlp. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics (ACL), 2020

  6. [6]

    Nuanced metrics for measuring unintended bias with real data for text classification

    Borkan, D., Dixon, L., Sorensen, J., Thain, N., and Vasserman, L. Nuanced metrics for measuring unintended bias with real data for text classification. In Companion proceedings of the 2019 world wide web conference, pp.\ 491--500, 2019

  7. [7]

    Language models are few-shot learners

    Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., et al. Language models are few-shot learners. In Advances in Neural Information Processing Systems (NeurIPS), 2020

  8. [8]

    S., Dai, H., and Dou, D

    Che, T., Liu, J., Zhou, Y., Ren, J., Zhou, J., Sheng, V. S., Dai, H., and Dou, D. Federated learning of large language models with parameter-efficient prompt tuning and adaptive optimization. arXiv preprint arXiv:2310.15080, 2023

Show all 35 references
  1. [9]

    Bert: Pre-training of deep bidirectional transformers for language understanding

    Devlin, J., Chang, M.-W., Lee, K., and Toutanova, K. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics (NAACL), 2019

  2. [10]

    Multi-dimensional gender bias classification

    Dinan, E., Golub, D., Wu, L., and Weston, J. Multi-dimensional gender bias classification. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), 2020

  3. [11]

    The llama 3 herd of models

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

  4. [12]

    Fairness through awareness

    Dwork, C., Hardt, M., Pitassi, T., Reingold, O., and Zemel, R. Fairness through awareness. In Proceedings of the 3rd innovations in theoretical computer science conference, pp.\ 214--226, 2012

  5. [13]

    User-item fairness tradeoffs in recommendations

    Greenwood, S., Chiniah, S., and Garg, N. User-item fairness tradeoffs in recommendations. arXiv preprint arXiv:2412.04466, 2024

  6. [14]

    Harper, F. M. and Konstan, J. A. The movielens datasets: History and context. ACM Transactions on Interactive Intelligent Systems (TiiS), 5 0 (4): 0 19:1--19:19, 2015

  7. [15]

    and McAuley, J

    He, R. and McAuley, J. Ups and downs: Modeling the visual evolution of fashion trends with one-class collaborative filtering. In Proceedings of the 25th International Conference on World Wide Web (WWW), pp.\ 507--517, 2016

  8. [16]

    Hou, Y., Zhang, J., Lin, Z., Lu, H., Xie, R., McAuley, J., and Zhao, W. X. Large language models are zero-shot rankers for recommender systems. In European Conference on Information Retrieval, pp.\ 364--381. Springer, 2024

  9. [17]

    Up5: Unbiased foundation model for fairness-aware recommendation

    Hua, W., Ge, Y., Xu, S., Ji, J., and Zhang, Y. Up5: Unbiased foundation model for fairness-aware recommendation. arXiv preprint arXiv:2305.12090, 2023

  10. [18]

    a rvelin, K. and Kek \

    J \"a rvelin, K. and Kek \"a l \"a inen, J. Cumulated gain-based evaluation of ir techniques. ACM Transactions on Information Systems (TOIS), 20 0 (4): 0 422--446, 2002

  11. [19]

    Q., Sablayrolles, A., Mensch, A., Bamford, C., Chaplot, D

    Jiang, A. Q., Sablayrolles, A., Mensch, A., Bamford, C., Chaplot, D. S., Casas, D. d. l., Bressand, F., Lengyel, G., Lample, G., Saulnier, L., et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023

  12. [20]

    and Bamman, D

    Lucy, L. and Bamman, D. Gender and representation bias in gpt-3 generated stories. In Proceedings of the 3rd Workshop on Narrative Understanding (NAACL), 2021

  13. [21]

    Learning adversarially fair and transferable representations

    Madras, D., Creager, E., Pitassi, T., and Zemel, R. Learning adversarially fair and transferable representations. In Proceedings of the 35th International Conference on Machine Learning (ICML), 2019

  14. [22]

    Image-based recommendations on styles and substitutes

    McAuley, J., Targett, C., Shi, Q., and van den Hengel, A. Image-based recommendations on styles and substitutes. In Proceedings of the 38th International ACM SIGIR Conference on Research and Development in Information Retrieval, pp.\ 43--52, 2015

  15. [23]

    Training language models to follow instructions with human feedback

    Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35: 0 27730--27744, 2022

  16. [24]

    Sentence-bert: Sentence embeddings using siamese bert-networks

    Reimers, N. Sentence-bert: Sentence embeddings using siamese bert-networks. arXiv preprint arXiv:1908.10084, 2019

  17. [25]

    and McDonell, K

    Reynolds, L. and McDonell, K. Prompt programming for large language models: Beyond the few-shot paradigm. arXiv preprint arXiv:2102.07350, 2021

  18. [26]

    and Vovk, V

    Shafer, G. and Vovk, V. A tutorial on conformal prediction. Journal of Machine Learning Research, 9 0 (3), 2008

  19. [27]

    Sharma, V., Mishra, N., Kukreja, V., Alkhayyat, A., and Elngar, A. A. Framework for evaluating ethics in ai. In 2023 International Conference on Innovative Data Communication Technologies and Application (ICIDCA), pp.\ 307--312. IEEE, 2023

  20. [28]

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

    Sheng, E., Chang, K.-W., Natarajan, P., and Peng, N. The woman worked as a babysitter: On biases in language generation. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing (EMNLP), 2019

  21. [29]

    Mitigating gender bias in natural language processing: Literature review

    Sun, T., Gaut, A., Tang, S., Huang, Y., Qian, H., Wang, S., and Choi, Y. Mitigating gender bias in natural language processing: Literature review. arXiv preprint arXiv:1906.08976, 2019

  22. [30]

    Llama 2: Open foundation and fine-tuned chat models

    Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi, A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava, P., Bhosale, S., et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023

  23. [31]

    Towards fairness in text classification: An overview of mitigation strategies

    Wang, L., Chen, L., Zhu, X., and Li, S. Towards fairness in text classification: An overview of mitigation strategies. ACM Computing Surveys, 55 0 (3), 2022

  24. [32]

    Let's erase: Prompt-based mitigation of textual bias

    Yang, Z., Zhang, Y., and Zhao, H. Let's erase: Prompt-based mitigation of textual bias. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (ACL), 2022

  25. [33]

    Is chatgpt fair for recommendation? evaluating fairness in large language model recommendation

    Zhang, J., Bao, K., Zhang, Y., Wang, W., Feng, F., and He, X. Is chatgpt fair for recommendation? evaluating fairness in large language model recommendation. In Proceedings of the 17th ACM Conference on Recommender Systems, pp.\ 993--999, 2023

  26. [34]

    Gender bias in coreference resolution: Evaluation and debiasing methods

    Zhao, J., Wang, T., Yatskar, M., Ordonez, V., and Chang, K.-W. Gender bias in coreference resolution: Evaluation and debiasing methods. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics (NAACL), 2018

  27. [35]

    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 gl...

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.