George Wang, Elizabeth Donoway, Daniel Murfetcs.LG
Reward models trained on human preferences are known to suffer from length, formatting, and other stylistic biases. In this paper we use patterning, which reweights each preference pair according to its measured effect on posterior expectation values of benchmark losses (its susceptibility), to debias a Gemma 2 9B Instruct reward model trained on Skywork-Reward-Preference v0.2. We obtain $+14.2 \pm 1.2$ pp on RM-Bench Hard, the split where style cues point against correctness (mean $\pm$ s.e.\ over 5 seeds), with overall RM-Bench accuracy preserved, comparable to the strongest Hard-split gain reported by the closest published comparator (SteerRM, $+13.2$ pp). We demonstrate in a simple case that the reweighting is interpretable by tracing a side effect of the intervention (a regression on a safety subset of RM-Bench) to a small class of training pairs, which we confirm by ablation. The weights also transfer: those computed on Gemma 2 9B debias Gemma 2 2B and 27B with no recomputation, and transfer partially to Llama 3.1 8B. This is the first application of patterning, a program grounded in singular learning theory, beyond small models and synthetic tasks.
Standard language proficiency tests rely on linguistic tasks such as vocabulary, grammar and reading comprehension quizzes. An alternative, cognitively motivated approach, introduced in Berzak et al. (2018), proposed instead to predict language proficiency from behavioral traces of eye movements in reading. In this work, we validate and extend this approach from single sentences to more naturalistic reading of contextualized passages in English as a second language, new proficiency measures, prediction models, and reading in an information seeking regime. We find that the approach is effective in all these evaluations. We further address two key open questions on eye movement based proficiency testing: (1) potential scoring biases that reflect the proximity of the reader's native language to English, which may undermine validity, and (2) its reliability. We find that eye movement based proficiency scores are indeed biased towards L1s that are linguistically closer to English. We propose a score debiasing method which effectively remedies this issue. The reliability analyses suggest that eye movement proficiency scores are more reliable than standard language proficiency scores. Overall, our results strengthen and broaden the empirical foundations for future eye movement based language assessment technologies.
Large Language Models (LLMs) often favor Western-associated entities across cultural contexts. Conventional debiasing methods aim for uniform neutrality, but cultural bias mitigation demands context-conditional behavior, preferring culturally appropriate entities when cultural cues are present and remaining neutral when they are absent. We propose CoCoA (Context-Conditional Cultural Alignment), a framework that learns this behavior through dual-context training on the same entity pairs under contexts with and without cultural cues. CoCoA combines a contrastive alignment objective with calibration and drift regularization, optimized through goal-aware gradient reconciliation. We evaluate CoCoA on CAMeL and Camellia, two entity-centric cultural bias benchmarks, across ten language settings and four LLMs. CoCoA reduces the Cultural Bias Score from 43 to 24 on average while maintaining near-neutral preferences at 50.2, with minimal impact on general performance across five standard benchmarks. These findings highlight that effective cultural alignment requires context-conditional modeling rather than uniform debiasing, and establish a new direction for mitigating entity-centric cultural bias in LLMs.
Generative vision-language models (VLMs) are increasingly used in human-centered settings, yet they can produce demographically biased outputs even when images differ only in controlled attributes such as perceived race or gender. However, existing inference-time debiasers were largely designed for static embeddings or CLIP-like models rather than generative VLMs. We propose GGSS---Geodesic-Gated Spherical Steering---a norm-preserving intervention that discovers a counterfactual bias subspace on the unit hypersphere, steers visual tokens along geodesic arcs, and uses an adaptive gate to focus correction on tokens that carry stronger demographic signal. We evaluate four generative VLMs against ten adapted inference-time debiasing baselines and prompt-based mitigation under a single operating-point protocol across categorical, pairwise, and occupation-gender bias tests, while also measuring general visual-language capability. GGSS achieves the lowest average bias on all four models, significant on three of four backbones under paired permutation tests, while preserving MMStar accuracy within +/- 0.6 p.p. of the unsteered baseline. Code is available at https://github.com/dukesun99/GGSS.
Multiple-choice benchmarks are widely used to evaluate large language models, but MCQ scores conflate knowledge with sensitivity to option order, which makes them unreliable measures of model knowledge. In this paper, we test whether preventing a model from seeing option labels while committing to an answer removes positional influence and, in turn, improves performance. We evaluate two different strategies for mitigating bias. The first uses a generation-then-matching approach, and the second scores options in isolation, which is positionally unbiased by construction. Neither reliably improves accuracy. A complete decomposition shows that the bottleneck is withholding options, not the matching step. The only configuration that consistently matches the baseline is the one that shows the model all options paired with an LLM matcher. However, eliminating positional influence entirely still does not reliably yield accuracy gains, while cyclic permutation often improves them. For two-stage prompting, an aggregate measure of recall imbalance and a direct per-question measure of order sensitivity both fail to show reliable debiasing.
Large vision-language models (LVLMs) have demonstrated strong performance in open-ended video understanding, yet they remain prone to fluent responses unsupported by video evidence. Existing training-free methods typically apply a globally fixed visual intervention or construct a contrastive branch through input perturbation. The former cannot accommodate video-dependent fusion paths, while the latter can be compensated by cross-frame redundancy. We therefore propose Video-Adaptive Debiasing via Evidence Reweighting (VADER), a training-free framework with two complementary modules. Visual Focus Reallocation (VFR) automatically instantiates an intervention policy for each video-question input: it diagnoses layer-wise visual-to-text evidence flow, determines where to intervene, and derives how strongly to reallocate pre-softmax attention from system-token to video-token blocks. Selective Evidence Erasure (SEE) independently masks high-importance visual tokens in every frame, constructing a prior-biased branch that is difficult to compensate through neighboring frames. Contrastive decoding then down-weights predictions that remain confident after selective evidence erasure. Across multiple VideoLLMs, VADER yields substantial improvements on event-level grounding and temporal consistency; on LLaVA-Video-7B, it reaches 72.60% accuracy on EventHallusion.
The visual reasoning ability of multimodal large language models (MLLMs) is crucial for downstream applications, particularly counter-commonsense reasoning, which requires models to reason beyond common assumptions. Recent studies mainly improve visual counter-commonsense reasoning by enhancing visual inputs, following the assumption that failures originate from insufficient visual grounding. However, our empirical analysis reveals that the bottleneck is not visual perception. MLLMs already capture the relevant visual evidence, and the correct answer exists in their decoding space. Instead, the shared language decoder resolves prior--evidence conflicts by favoring dominant language priors, especially for low-frequency factual scenarios. Motivated by this, we first propose a text-anchored data construction pipeline, whose core component, Fact-Frequency Distillation (FFD), estimates the prior strength of commonsense facts and distills verified counter-commonsense scenarios into a high-quality text corpus. Building upon this corpus, we introduce TACT, a text-anchored post-training framework that debiases the shared language decoder without requiring any visual training data. TACT routes evidence-following and prior-driven reasoning trajectories into different optimization stages, enabling the decoder to resolve prior--evidence conflicts. Across counter-commonsense visual benchmarks, TACT substantially improves visual reasoning while preserving general capabilities, demonstrating effective text-to-vision cross-modal transfer.
Language models (LMs) often acquire various biases during pre-training and may express them in interactions, potentially causing social harm. Existing methods often rely on counterfactual augmentation or representation projection. These strategies remain limited in practice due to their high computational costs and difficulty in scaling to larger models. Additionally, many of these strategies require manual data annotation, narrowing their scope to specific cultures and bias categories. To overcome these limitations, we propose HEIMAT, a HEurIstic-style autoMATic debiasing framework for LMs. HEIMAT consists of two main steps: bias disclosure and debiasing fine-tuning. In the first step, it uses simple templates to construct heuristic prompts, which are applied to reveal model biases and generate corresponding context prompts. In the second step, it fine-tunes the model by minimizing the Jensen-Shannon divergence of predictions on these context prompts to reduce bias. Extensive experiments show that HEIMAT effectively mitigates bias in different cultures while maintaining the model's natural language understanding (NLU) performance.
Olafur Gudmundsson, Bo Zhao, Huayi Liao +15cs.IR cs.LG
In this paper, we propose a new solution for addressing the content cold-start problem in industry-scale search and recommender systems. Compared to prior approaches, we have made the following new contributions: 1) our solution spans the entire multi-stage funnel and generalizes well for both search and recommendation surfaces, 2) our solution reduces bias favoring existing content, allowing more accurate model prediction across content types and reducing short-term tradeoffs associated with high volumes of explicit content exploration, 3) our solution is evaluated with a scalable measurement framework that enables fast short-term experimentation while validating long-term impact. We have iteratively built and successfully deployed this new system at Pinterest in the past two years and observed significant improvements in fresh content exploration, overall user engagement, and content ecosystem health.
Consider the partial linear model $Y = μ_0(X) + β_0 \cdot T + \varepsilon$ and $T = π_0(X) + u$ in the structure-agnostic setting, where we are blind to the structure $μ_0$ and $π_0$ and estimate the nuisances by a black-box hypothesis class. The learnability of the class is characterized by the estimation error $δ_s$ in the absence of model misspecification and its $L_2$ mis-specification error $δ_{a, μ}$ and $δ_{a, π}$ for $μ_0$ and $π_0$, respectively. We propose a novel estimator of the target linear coefficient $θ_0 = β_0$ with error rate \[ \frac{1}{\sqrt{n}} + δ_{a, μ} \cdot δ_{a, π} + [δ_s]^2. \] A matching lower bound is also established, implying that this rate is unimprovable. Compared with the product rate yielded by double machine learning (DML), our estimator removes the suboptimal term $\max(δ_{a, μ}, δ_{a, π})\cdot δ_s$ at no extra cost or assumption. Building on the underlying insights, which are neither tailored to the one-learner setting nor the partial linear model, we propose Transductive Adversarial Moment-calibrated Editing (TAME), which locally edits debiasing weights induced by black-box regression estimates on the inference sample through adversarial conditional moment calibration. TAME can be combined with any initial black-box estimates and can strictly improve on DML guarantees when the nuisance difficulties are imbalanced. We discuss how to fully exploit the advantages introduced by TAME, including the gains from using two learners, the resulting under-smoothing principle for model selection, and extensions to other linear functional estimation problems.
Prakhar Gupta, Terry Jingchen Zhang, Florent Draye +2cs.CL cs.AI cs.LG
Modern LLMs are alarmingly susceptible to surprisingly simple immaterial changes of input prompts: a casual hint, an incorrectly labeled few-shot example, or a fake prior assistant turn often flips an originally correct answer. We study where this susceptibility, spanning sycophancy and related cue-induced biases, lives inside the model. Across five model families and seven bias types, we extract a per-bias direction from hidden states and triangulate it through three measures: probing, leave-one-dataset-out (LODO) transfer, and causal intervention. The susceptibility is largely shaped by alignment tuning rather than pretraining: pretrained base models generally cave much less to these biases, and their activations carry much weaker cue-specific signal beyond question content. Within aligned models, each bias has a coherent linear direction that we can both decode and steer along, recovering the unbiased answer across every family we test. The biases do not collapse into a single shared representation, however: cross-bias overlap is model-specific, and even behaviorally similar biases occupy different directions. The same intervention also provides a proof-of-concept debiasing tool, recovering a meaningful share of bias-induced errors while preserving most correct answers across all instruct families. Cue-induced bias is therefore best understood not as a single flaw in LLMs but as a family of causally effective linear directions that are largely shaped by alignment tuning.
Yahan Zheng, John Guerrerio, Soroush Vosoughi +1cs.CL
Preprocessing-based methods for stereotype mitigation, such as pre-/post-training on debiased corpora, are widely used in NLP. While these approaches reduce measurable stereotypes for targeted groups, we find they often induce unintended shifts-side effects, where stereotyping or counter-stereotyping can increase relative to neutral baselines for other demographics, including across unrelated demographic categories. We demonstrate these side effects across two model families (encoder-only and decoder-only), multiple preprocessing strategies (removing stereotypical sentences, removing group mentions, and swapping group references), and both pre- and post-training at different data scales on Wikipedia. Standard benchmarks frequently miss these shifts. Using attention-rollout analysis, we observe that such side effects are not accompanied by large changes in attention flow, complicating mechanistic explanations. We discuss implications for evaluation, provide actionable diagnostics, and argue for side-effect-aware, transparent mitigation practices.
Text-to-image diffusion models power everyday creative tasks, but they still reproduce the demographic biases in their training data. On common prompts such as ``a photo of a nurse,'' ``a photo of a CEO'', they skew their outputs toward one gender, driven by the statistics of training data rather than anything in the text. Existing debiasing methods show promise in narrow settings but require retraining, batch-level control, or prompt-specific tuning, limiting their scalability. We propose \emph{EquiSteer}, a training-free method that works per sample by steering cross-attention (CA) activations at inference time. For each target attribute, EquiSteer precomputes steering vectors from contrastive prompts. Then at generation time, a prompt-aware gate leaves attribute-specific prompts untouched, while for neutral ones it clears existing attribute signals from the CA activations and injects a target attribute. Across SD-1.5, SD-2.1, SDXL, and SANA, EquiSteer reduces the average parity gap by up to $87\%$, with minimal effect on image quality and text-image alignment. Code is available at \href{https://github.com/Atmyre/EquiSteer}{https://github.com/Atmyre/EquiSteer}.%
Text-to-image diffusion models achieve impressive visual quality, yet demographic bias remains a challenge, as neutral prompts consistently produce stereotypical representations across gender and race. Existing approaches remain limited by costly retraining or by inference-time interventions that often degrade image quality and semantic alignment. We propose Text Embedding Steering (TES), a training-free framework that mitigates demographic bias by directly optimizing conditional text embeddings during the diffusion process. We show that a two-stage strategy - early-stage global alignment followed by iterative denoising-time refinement with CLIP-based feedback - enables stable and controllable attribute steering without modifying model parameters. Extensive experiments on Stable Diffusion demonstrate that TES outperforms existing training-free baselines in fairness while maintaining competitive image quality. These results highlight that inference-time text embedding optimization is a practical and scalable solution for fairness-aware generation in diffusion models.
Vision language models (VLMs) demonstrate strong zero-shot performance, but often perpetuate social stereotypes in person-centric queries, yielding skewed demographic distributions. Current debiasing methods apply uniform bias corrections across all input queries regardless of their bias sensitivity, creating a fundamental fairness--utility trade-off. Strong debiasing distorts semantically meaningful information in bias-insensitive queries, while weak debiasing fails to mitigate stereotypes in bias-sensitive ones. This one-size-fits-all approach hampers simultaneously achieving high utility on bias-insensitive queries and fairness on bias-sensitive queries. We introduce Reward-Gated Test-Time Adaptation (RG-TTA), a reinforcement learning-based test-time adaptation framework that selectively applies debiasing based on input sensitivity. RG-TTA adaptively triggers fairness regularization based on the bias sensitivity of each input during test-time policy adaptation, while focusing exclusively on optimizing cross-modal alignment for bias-insensitive inputs. Experiments on fairness benchmarks (e.g., FairFace, UTKFace) demonstrate substantial bias reduction while simultaneously improving zero-shot utility, resolving the trade-off of uniform debiasing.
Foundation models trained on biased datasets often rely on spurious correlations between target labels and non-causal attributes, resulting in poor generalization on minority groups. Bias mitigation remains challenging due to two fundamental issues. First, when group labels are unavailable, existing group-unsupervised methods typically infer spurious attributes implicitly from model behavior, making it difficult to identify spurious factors that are semantically aligned with real-world biases. Second, even with pseudo spurious supervision, most existing debiasing methods follow a single-branch design that operates within a single shared feature space, where target and spurious attributes are intrinsically entangled. To address the first challenge, we introduce Confidence-guided Bias Concept Mining (CBCM), which leverages diffusion-disentangled, semantically grounded concept representations to identify reliable spurious attributes without attribute annotations. To address the second challenge, we propose Dual-branch Cross-projection Debiasing (DCD), a prompt-tuning framework that separates target and spurious representations into two branches and explicitly removes spurious information through cross null-space projection while preserving target-relevant semantics. Extensive experiments on four benchmark datasets show that our method achieves state-of-the-art worst group accuracy among group-unsupervised approaches, while tuning at most 0.22% of the model parameters. The source code is available in the supplementary materials.
Benchmark-based evaluation is the dominant paradigm for assessing large language model (LLM) capabilities, yet data contamination inflates reported performance and undermines fair comparison. Existing decontamination methods are evaluated solely through aggregate accuracy, which can obscure substantial differences in per-sample model behaviour, and many require access to an uncontaminated model. In this paper, we propose a sample-level evaluation framework for decontamination that complements accuracy-based assessment with distributional distance metrics, measuring how closely a decontaminated model recovers the output distribution of an uncontaminated model on each sample. Building on this framework, we introduce Uncertainty-Based Decontamination (UBD), a family of methods that leverage deep ensembles of the contaminated model to estimate per-sample memorization without requiring a uncontaminated model or knowledge of which samples are contaminated. UBD estimates a per-sample correction scalar from ensemble uncertainty, which is used to construct a debiased target distribution that suppresses the inflated probability mass on correct answers induced by contamination. This target is then used either as a post-hoc output correction (debiasing) or as a soft training signal for parameter update (unlearning). Experiments on MMLU-Pro and MATH-MCQA across multiple LLM backbones demonstrate that UBD produces per-sample output distributions substantially closer to those of an uncontaminated model than paraphrasing or choice-permutation baselines, while preserving model performance on uncontaminated data.
Most fairness research in NLP assumes direct access to protected attributes such as gender, race, or nationality. In practice, however, such information is often unavailable due to privacy constraints, missing metadata, or legal restrictions, even though models may infer it from indirect textual cues. This raises a key question: can debiasing succeed without direct access to sensitive attributes? We propose H-SAL, which performs post-hoc concept and attribute erasure using self-description text as an implicit debiasing signal. To support this setting, we introduce a multi-domain Stack Exchange-based fairness benchmark for helpfulness prediction that includes both explicit and implicit signals, enabling comparison between standard debiasing with protected labels and debiasing without access to sensitive information. Across encoder and decoder-only language models, we find that implicit self-description often matches or outperforms explicit-label-based debiasing. Our results broaden representation-level fairness research and provide a new benchmark for studying debiasing under realistic data constraints.
Alexey Kresin, Tchifou M. Dieffi, Tomer Caspics.CL
Debiasing methods based on principal component analysis (PCA) are broadly used to reduce gender bias in word embeddings used in LLMs, yet it remains unclear what aspects of bias they actually remove and how destructive this process is. These methods are based on the understanding that bias resides in a low-dimensional subspace, with the assumption that most of it can be captured by a few principal components. In this work, we conduct a systematic geometric analysis of PCA-based gender debiasing and investigate what is actually removed from the embedding space. Our experiments across multiple embeddings show that direct gender bias is primarily concentrated in the first principal component, supporting the low-rank bias hypothesis. However, associative bias measured by WEAT does not align with these principal directions and is instead spread across multiple embedding dimensions. Furthermore, as expected, we demonstrate that removing an increasing number of principal components leads to a consistent degradation of the embedding geometry, affecting semantic structure and vector relationships. These results reveal that PCA-based debiasing operates as a trade-off: while it effectively reduces certain forms of direct bias, it fails to eliminate distributed associations and introduces geometric distortion. Moreover, there is no universal optimal level of debiasing, as the balance between bias reduction and semantic preservation depends on the chosen metric and embedding. Overall, our findings suggest that bias in word embeddings is not purely low-rank and that simple subspace removal methods may be insufficient for comprehensive debiasing.
This paper explores policy learning from observational data, focusing on a nonlinear welfare criterion in a binary treatment setting. The nonlinear criterion is inspired by scenarios where policymakers prioritize specific population segments. We model this criterion using a utility function that encompasses potential outcomes and intermediate parameters, with the latter capturing higher moments of the outcome distributions. When formulated in the context of observational data, both the intermediate parameters and the welfare criterion depend on the propensity score, which we estimate using machine-learning techniques. To address bias in machine learning estimates, we introduce a novel reweighting-based debiasing approach that offers a promising alternative to traditional orthogonality-based methods. To tackle the complexities of infinite-dimensional policy spaces, we employ sieve approximations and $K$-fold cross-validation for model selection, thereby fully automating the policy-learning process. Despite these complexities, we demonstrate that both the welfare regret and the average welfare regret of our proposed policy learning method satisfy an oracle inequality, thereby providing theoretical guarantees on the performance of the estimated policy relative to the best possible policy. This finding extends the existing results from linear to nonlinear welfare criteria, from finite-dimensional to infinite-dimensional policy spaces, and from a known propensity score to a machine-learned one.
Fares El Khoury, Houssam Zenati, Nathan Kallus +2stat.ML cs.LG
Functional bilevel methods estimate a lower-level function and plug it into a hypergradient, but this plug-in gradient can retain first-order bias when the lower-level problem is learned nonparametrically. To remove this bias, we develop a semiparametric debiasing theory for population bilevel gradients based on the efficient influence function. This perspective leads to a cross-fitted orthogonal hypergradient estimator for which we establish asymptotic normality together with uniform control over the outer parameter. Under quadratic losses, the estimator reduces to a simple doubly robust score based on conditional mean nuisances. On synthetic bilevel benchmarks with known ground truth, the method tracks the oracle efficient-gradient benchmark and improves over plug-in functional hypergradients and regularized kernel bilevel baselines.
Automated glaucoma detection is critical for preventing irreversible vision loss and reducing the burden on healthcare systems. However, ensuring fairness across diverse patient populations remains a significant challenge. In this paper, we propose FairEnc, a fair pretraining method for vision-language models (VLMs) that enables simultaneous debiasing across multiple sensitive attributes. FairEnc jointly mitigates biases in both textual and visual modalities with respect to multiple sensitive attributes, including race, gender, ethnicity, and language. Specifically, for the textual encoder, we leverage a large language model to generate synthetic clinical descriptions with varied sensitive attributes while preserving disease semantics, and employ a contrastive alignment objective to encourage demographic-invariant representations. For the visual encoder, we propose a dual-level fairness strategy that combines mutual information regularization to reduce statistical dependence between learned features and demographic groups, with multi-discriminator adversarial debiasing. Comprehensive experiments on the publicly available Harvard-FairVLMed dataset demonstrate that FairEnc effectively reduces demographic disparity as measured by DPD and DEOdds while achieving strong diagnostic performance under both zero-shot and linear probing evaluations. Additional experiments on the private FairFundus dataset show that FairEnc consistently preserves fairness advantages under cross-domain and cross-modality settings and maintains diagnostic performance within a competitive range. These results highlight FairEnc's ability to generalize fairness under distribution shifts, supporting its potential for more equitable deployment in real-world clinical settings. Our codebase and synthetic clinical notes are available at https://github.com/Mohamed-Elhabebe/FairEnc
Reward models (RMs) play a central role in aligning large language models (LLMs) with human preferences. However, RMs are often sensitive to spurious features such as response length. Existing inference-time approaches for mitigating these biases typically focus exclusively on response length, resulting in performance trade-offs. In this paper, we propose causally motivated intervention for mitigating multiple types of biases in RMs at inference time. Our method first identifies neurons whose activations are strongly correlated with predefined bias attributes, and applies neuron-level intervention that suppresses these signals. We evaluate our method on RM benchmarks and observe reductions in sensitivity to spurious features across diverse bias types, without inducing performance trade-offs. Moreover, when used for preference annotation, small RMs (2B and 7B) with our method, which edits less than 2% of all the neurons in RMs, enable LLMs to improve alignment, achieving performance comparable to that of a state-of-the-art 70B RM on AlpacaEval and MT-Bench. Further analysis reveals that bias signals are primarily encoded by neurons in early layers, shedding light on the internal mechanisms of bias exploitation in RMs.