Bias Mitigation

Large language models (LLMs) have emerged as a standard paradigm for automated multilingual evaluation, yet exhibit systematic biases. In this paper, we identify ``translationese bias'', in which LLMs systematically favor machine-translated text over human-authored references, and this bias is particularly pronounced in low-resource languages. We attribute this bias to spurious correlations with (a) strong latent manifold isomorphism with English and (b) high predictive confidence. To mitigate these issues, we present DIBJudge, a robust fine-tuning framework that decouples robust features and bias representations by explicitly isolating spurious attributes into a dedicated bias branch and penalizing mutual dependence to enforce disentanglement. In particular, we present a vector-quantized compression that ensures the robust representation retains minimal and sufficient judgment-critical information. Extensive evaluations on multilingual reward modeling benchmarks and a specially designed translationese bias evaluation suite demonstrate that DIBJudge outperforms strong baselines and effectively mitigates translationese bias.

Multimodal Large Language Models (MLLMs) have shown strong performance in multi-image cross-modal retrieval, yet suffer from severe position bias, where predictions are dominated by input order rather than semantic relevance. Through empirical analysis, we identify a phenomenon termed Logit-Attention Divergence, in which output logits are heavily biased while internal attention maps remain well-aligned with relevant visual evidence. This observation reveals a fundamental limitation of existing logit-level calibration methods such as PriDe. Based on this insight, we propose a training-free, attention-guided debiasing framework that leverages intrinsic attention signals for instance-level correction at inference time, requiring only a minimal calibration set with negligible computational overhead. Experiments on MS-COCO-based benchmarks show that our method substantially improves permutation invariance and achieves state-of-the-art performance, enhancing accuracy by over 40\% compared to baselines. Code is available at https://github.com/brightXian/LAD.

Watermarking offers a promising solution for detecting LLM-generated content, yet its robustness under realistic query-free (black-box) evasion remains an open challenge. Existing query-free attacks often achieve limited success or severely distort semantic meaning. We bridge this gap by theoretically analyzing rewriting-based evasion, demonstrating that reducing the average conditional probability of sampling green tokens by a small margin causes the detection probability to decay exponentially. Guided by this insight, we propose the \emph{Bias-Inversion Rewriting Attack} (BIRA), a practical query-free method that applies a negative logit bias to a proxy suppression set identified via token surprisal. Empirically, BIRA achieves state-of-the-art evasion rates ($>99\%$) across diverse watermarking schemes while preserving semantic fidelity substantially better than prior baselines. Our findings reveal a fundamental vulnerability in current watermarking methods and highlight the need for rigorous stress tests. Our code is available at \href{https://github.com/ml-postech/LLM-Watermark-Evasion-via-Bias-Inversion}{here}.

Differentially private (DP) training protects individual examples by adding noise to gradients, but the injected noise interacts nontrivially with adaptive optimizers. Recent DP methods temporally filter privatized gradients to reduce variance; however, filtering also changes the DP noise statistics seen by AdamW’s second-moment accumulator. As a result, bias corrections derived for unfiltered DP noise (e.g., subtracting $\sigma_w^2$) can become miscalibrated when filtering is present. We propose FIBER, a DP optimizer designed for temporally filtered privatized gradients. FIBER (i) performs denoising in innovation space by filtering the residual stream and integrating it to form the filtered gradient estimate, (ii) decouples the two-point observation geometry from the innovation gain to enable independent tuning, and (iii) introduces a filter-aware second-moment calibration that subtracts the attenuated DP noise contribution $A(\omega)\sigma_w^2$, where $A(\omega)$ is derived in closed form for the innovation filter and can be computed for general stable linear filters. Across vision and language benchmarks, FIBER consistently demonstrates substantial improvements in the performance of DP optimizers, surpassing state-of-the-art results under equivalent privacy constraints on multiple tasks.

Federated Learning (FL) often suffers from degraded generalization under statistical heterogeneity, where client updates systematically deviate from the global objective. While recent Sharpness-Aware Minimization (SAM) methods promote locally flat solutions, they implicitly assume that local flatness transfers to the global model, which generally does not hold under heterogeneous data distributions. This mismatch gives rise to a flatness discrepancy induced by misaligned loss landscapes. To address this issue, we propose FedScar, a federated optimization framework that explicitly corrects heterogeneityinduced geometric inconsistency. FedScar maintains a history-accumulated geometric bias to capture persistent curvature skew across clients, and employs a variance-aware injection mechanism to steer local updates toward regions that are flat with respect to the global objective. We provide a theoretical interpretation of FedScar as a Split-Dual ADMM formulation, which jointly enforces parameter consensus and geometric alignment. Extensive experiments under severe heterogeneity demonstrate that FedScar consistently reduces flatness discrepancy and improves generalization over state-of-the-art methods, without incurring additional communication overhead. The source code is available at https://github.com/Ultraman6/FedScar.

Weak-to-strong generalization (W2SG) is the phenomenon in which a powerful student model, trained on labels produced by a weaker teacher, ultimately outperforms the teacher on the target task. In this work, we theoretically investigate how W2SG can arise via a generalized bias–variance decomposition under Bregman divergence. We show that the expected population risk gap between the student and the teacher is characterized by the expected misfit between the two models. Unlike earlier misfit-based analyses, our theory removes several restrictive assumptions, e.g., it does not require the student hypothesis class to be convex. Our results indicate that W2SG is more likely when the student effectively approximates the teacher's posterior mean. Specializing to squared loss, we provide a sufficient condition (illustrated through a concrete example) under which the student converges to its posterior mean teacher; in particular, increasing the student model size can ensure this convergence. For cross-entropy loss, our analysis further suggests that lowering the entropy of the student's predictive distribution can promote W2SG. We also find that the reverse cross-entropy, unlike the standard forward cross-entropy, is less sensitive to the teacher's predictive uncertainty. Finally, we verify these theoretical insights empirically and demonstrate that incorporating reverse cross-entropy consistently improves student performance.

Autoregressive large language models (LLMs) have achieved remarkable success in many complex tasks, yet they can still fail in very simple logical reasoning such as the "reversal curse" --- when trained on forward knowledge data of the form "$A \rightarrow B$" (e.g., Alice's husband is Bob), the model is unable to deduce the reversal knowledge "$B \leftarrow A$" (e.g., Bob's wife is Alice) during test. Extensive prior research suggests that this failure is an inherent, fundamental limit of autoregressive causal LLMs, indicating that these models tend to memorize factual-level knowledge rather than capture higher-level rules. In this paper, we challenge this view by showing that this seemingly fundamental limit can be mitigated by slightly tweaking the training data with a simple regularization data recipe called the Identity Bridge of the form "$A \to A$" (e.g., The name of Alice is Alice). Theoretically, we prove that under this recipe, even a one-layer transformer can break the reversal curse by analyzing the implicit bias of gradient descent. Empirically, we show that a 1B pretrained language model finetuned with the proposed data recipe achieves a 50\% success rate on reversal tasks, in stark contrast to a near-zero success rate when trained solely on forward-knowledge data. Our work provides a novel theoretical foundation for the reversal curse and offers a principled, low-cost path to encouraging LLMs to learn higher-level rules from data.

Despite their impressive performance on a wide variety of tasks, modern language models remain susceptible to distribution shifts, exhibiting brittle behavior when evaluated on data that differs in distribution from their training data. In this paper, we describe how distribution shifts in language models can be separated into *observable* and *unobservable* components, and we discuss how established approaches for dealing with distribution shift address only the former. Importantly, we identify that the resulting *omitted variable bias* from unobserved variables can compromise both evaluation and optimization in language models. To address this challenge, we introduce a framework that maps the strength of the omitted variables to bounds on the *worst-case generalization performance* of language models under distribution shift. In empirical experiments, we show that using these bounds directly in language model evaluation and optimization provides more principled measures of out-of-distribution performance, improves true out-of-distribution performance relative to standard distribution shift adjustment methods, and further enables inference about the strength of the omitted variables when target distribution labels are available.

Large vision language models (LVLMs) have made rapid advancements and are deployed across various applications, yet hallucinations remain a major challenge. Activation steering is appealing due to its minimal training overhead and controllability at inference time. However we found that during autoregressive decoding, visual conditioning affects token prediction sparsely and locally across decoding steps, and many existing methods that average image-versus-no-image differences over the entire sequence dilute these critical signals, yielding low signal-to-noise ratio steering directions. Additionally, many existing methods apply a fixed steering strength, which misallocates the intervention budget, over-perturbs non-critical tokens, and can cause instability. To address these limitations, we propose Token-Level Visual-Sensitivity Steering (TLVS) for hallucination mitigation. Our approach first extract token-level steering vectors and refines them, and then applies fine-grained, visual-sensitivity–adaptive steering only where it matters. This lightweight, plug-and-play mechanism requires only minimal training for calibration and can be applied across diverse vision-language models. It modulates the steering strength at each decoding step, selectively suppressing hallucination-prone spans while preserving evidence-grounded content. We evaluate TLVS on several benchmarks, including POPE, AMBER, CHAIR (COCO), MMHal and HallusionBench, demonstrating consistent improvements over previous steering methods.

Time series forecasting plays a vital role in numerous domains. However, real-world time series are frequently contaminated by noise, missing values, and anomalies, posing significant challenges to reliable forecasting. In this work, we first systematically investigate a fundamental limitation prevalent in existing forecasting methods: an excessive reliance on the most recent observations---termed "recent-data bias". This bias renders forecasts highly vulnerable to perturbations in recent data, severely undermining prediction reliability. To address this issue, we propose TameR, a novel approach for robust time series forecasting that effectively mitigates recent-data bias via enhancing the utilization of global context. Specifically, it employs a basis-aligned randomized sampling strategy to reduce dependence on any specific recent data. Furthermore, TameR incorporates a learnable periodicity extraction module coupled with a two-stage learning protocol to robustly separate periodic patterns from the sampled residual components. Comprehensive experiments demonstrate that TameR significantly outperforms state-of-the-art methods in robustness against diverse perturbation scenarios, while achieving comparable accuracy on clean data. Code is available at https://github.com/NetManAIOps/TameR.

Although Large Language Models (LLMs) demonstrate remarkable reasoning capabilities, inherent social biases often cascade throughout the Chain-of-Thought (CoT) process, leading to continuous "Bias Propagation". Existing debiasing methods primarily focus on static constraints or external interventions, failing to identify and interrupt this propagation once triggered. To address this limitation, we introduce Self-Debias, a progressive framework designed to instill intrinsic self-correction capabilities. Specifically, we reformulate the debiasing process as a strategic resource redistribution problem, treating the model's output probability mass as a limited resource to be reallocated from biased heuristics to unbiased reasoning paths. Unlike standard preference optimization which applies broad penalties, Self-Debias employs a fine-grained trajectory-level objective subject to dynamic debiasing constraints. This enables the model to selectively revise biased reasoning suffixes while preserving valid contextual prefixes. Furthermore, we integrate an online self-improvement mechanism utilizing consistency filtering to autonomously synthesize supervision signals. With merely 20k annotated samples, Self-Debias activates efficient self-correction, achieving superior debiasing performance while preserving general reasoning capabilities without continuous external oversight.

Dataset bias often leads deep learning models to exploit spurious correlations instead of task-relevant signals. We introduce the Standard Anti-Causal Model (SAM), a unifying causal framework that characterizes bias mechanisms and yields a conditional independence criterion for causal stability. Building on this theory, we propose DISCO$_m$ and sDISCO, efficient and scalable estimators of conditional distance correlation that enable independence regularization in gradient-based models. Across six diverse datasets, our methods consistently outperform or are competitive in existing observed bias mitigation approaches, while requiring fewer hyperparameters and scaling seamlessly to multi-bias scenarios. This work bridges causal theory and practical deep learning, providing both a principled foundation and effective tools for robust prediction. Source Code: https://github.com/yakamoz5/DISCO.

Neural collapse (NC) describes the structured geometry that emerges in the features and weights of trained classifiers. Recent theory suggests NC can be suboptimal in deep architectures, attributing this to an explicit low-rank bias from L2 regularization. We study the deep unconstrained feature model (UFM)—equivalent to a deep linear network with orthogonal inputs—trained without regularization, to isolate how gradient descent and depth alone shape NC. We show that depth induces an implicit low-rank bias: low-rank matrices propagate norm more efficiently through successive multiplications, promoting low-rank alternatives to NC. These alternatives, we argue, correspond to softmax codes: max-margin solutions previously found in width-bottlenecked networks. Analyzing training dynamics under spectral initialization, we identify an early-time repulsion among singular values that drives low-rank emergence, and characterize how depth shrinks NC's basin of attraction. Finally, we show that some effects act in the opposite direction: for randomly initialized networks, increasing width biases training toward higher-rank solutions. Our results provide the first asymptotic and dynamic characterization of implicit bias in deep UFMs trained with unregularized multiclass cross-entropy.

We generalize the attention mechanism by viewing it through the lens of Entropic Optimal Transport, revealing that standard attention corresponds to a transport problem regularized by an implicit uniform prior. We introduce Generalized Optimal transport Attention with Trainable priors (GOAT), a new attention mechanism that replaces this naive assumption with a learnable, continuous prior. This prior maintains full compatibility with optimized kernels such as FlashAttention. GOAT also provides an EOT-based explanation of attention sinks and materializes a solution for them, avoiding the representational trade-offs of standard attention. Finally, by absorbing spatial information into the core attention computation, GOAT learns an extrapolatable prior that combines the flexibility of learned positional embeddings with the length generalization of fixed encodings.

A variety of widely used optimization methods like SignSGD and Muon can be interpreted as instances of steepest descent under different norm-induced geometries. In this work, we study the implicit bias of mini-batch stochastic steepest descent in multi-class classification, characterizing how batch size, momentum, and variance reduction shape the limiting max-margin behavior and convergence rates under general entry-wise and Schatten-$p$ norms. We show that, without momentum, worst-case convergence and successful classification can only be guaranteed with full-batch gradient. In contrast, momentum enables small-batch convergence to an approximate max-margin solution through a batch-momentum trade-off, though it slows convergence. This approach provides fully explicit, dimension-free rates that improve upon prior results. Moreover, we prove that variance reduction can recover the exact full-batch implicit bias for any batch size, albeit at a slower convergence rate. Finally, we further investigate the batch-size-one steepest descent without momentum, and reveal its convergence to a fundamentally different bias via a concrete data example, which reveals a key limitation of purely stochastic updates. Overall, our unified analysis clarifies when stochastic optimization aligns with full‑batch behavior, and paves the way for perform deeper explorations of the training behavior of stochastic gradient steepest descent algorithms.

Vision language models (VLMs) often generate hallucination, i.e., content that cannot be substantiated by either textual or visual inputs. Prior work primarily attributes this to over-reliance on linguistic prior knowledge rather than visual inputs. Some methods attempt to mitigate hallucination by amplifying visual token attention proportionally to their attention scores. However, these methods overlook the visual attention sink problem, where attention is frequently misallocated to task-irrelevant visual regions, and neglect cross-modal fusion balance by enhancing only visual attention without adjusting attention to the user query. This can result in amplifying incorrect areas while failing to properly interpret the user query. To address these challenges, we propose a simple yet effective method called Gaze Shift-Guided Cross-modal Fusion Enhancement (GIFT). GIFT pre-computes a holistic visual saliency map by tracking positive changes in visual attention, or "gaze shifts", during user query comprehension, and leverages this map to amplify attention to both salient visual information and the user query at each decoding step. This reduces the impact of visual attention sink, as irrelevant tokens exhibit minimal shifts, while ensuring balanced cross-modal fusion for well-integrated representation. Extensive experiments show that GIFT effectively mitigates hallucination in VLMs across both generative and classification tasks, achieving up to 20.7% improvement over greedy decoding, while maintaining general vision-language performance with low computational overhead.

Generations from large language models often fail to conform to desired constraints such as JSON schema. Existing locally constrained decoding (LCD) approaches enforce constraints by myopically masking out next tokens, resulting in biased sampling and degradation in performance. Recent work uses sequential Monte Carlo (SMC) methods to mitigate such biases, but designing effective proposal distributions or potential functions remains a key challenge. In this work, we propose a generic approach to construct proposals and potentials for SMC sampling from $p_{\mathrm{lm}}( \cdot \mid \mathrm{constraint})$. First, we show that constraints specified as finite automata can be tensorized for efficient execution on GPUs, which we use to construct globally constrained decoding (GCD) proposals. In addition, leveraging the fact that tensorized finite automata share the same circuit structure as hidden Markov models, we circuit-multiply them to obtain the probabilistic GCD (P-GCD) proposals encoding both logical and probabilistic information about the target distributions. We evaluate (P-)GCD on the tasks of function calling, keyword-based generation, and SQL generation. Experiments show that under the same SMC sampling setup, compared to LCD proposals, (P-)GCD converges faster to the target distribution with significantly fewer particles.

Zeroth-order optimization is widely used when gradients are unavailable, but the standard two-point estimator suffers from $\mathcal{O}(r^2)$ truncation bias at smoothing radius $r$. Existing bias-reduction schemes typically increase the leading directional variance under a fixed number of function evaluations per gradient estimate, while variance-reduction schemes generally do not improve the bias order. We propose Romberg-ZOGE, which forms a Romberg-extrapolated linear combination of two-point differences evaluated at radii $\{r/2^k\}_{k=0}^R$ while reusing the same perturbation direction across all radii. With appropriately chosen weights, Romberg-ZOGE cancels the first $R$ even-order truncation terms and achieves $\mathcal{O}(r^{2R+2})$ bias under $(2R{+}2)$-order smoothness, while preserving the leading directional variance constant of the two-point estimator up to higher-order residual terms. We further characterize the stochastic-oracle setting by deriving an explicit noise-amplification factor and corresponding bias and variance bounds. Experiments on synthetic benchmarks, simulator-based wireless optimization, and black-box prompt tuning of OPT-1.3B demonstrate faster and more stable zeroth-order SGD when the number of function evaluations per gradient estimate is fixed.

The proliferation of recursive training on synthetic data can alleviate data scarcity but risks model collapse, where repeated training erodes distributional tails and homogenizes outputs. Data selection is widely viewed as a remedy, yet its reliability depends critically on the reference distribution used by the verifier. We show that in low-resource verification regimes, where each verifier observes only a small, fragmented, and biased slice of the target manifold, selection itself becomes biased. This situation naturally arises in low-resource data silos such as healthcare consortia or proprietary financial institutions, where raw data cannot be pooled and local references are inherently incomplete. As a result, selection preferentially retains samples aligned with the local manifold while pruning globally relevant tail modes, turning from a safeguard against collapse into a mechanism that precipitates it. We theoretically prove that such siloed selection accelerates collapse and induces power-law diversity decay. As an initial mitigation, we construct Wasserstein proxy references from multiple silos without sharing raw data. Empirical results confirm that local-reference selection fails on skewed distributions, whereas collaborative proxy references mitigate diversity degradation, suggesting that recursive synthetic-data pipelines require particular caution when real-data coverage is fragmented or scarce.

We study the implicit bias of momentum-based optimizers on smooth homogeneous models. We show that *momentum steepest descent* algorithms like Muon (spectral norm), MomentumGD ($\ell_2$ norm), and Signum ($\ell_\infty$ norm) are *approximate* steepest descent trajectories under a decaying learning rate schedule, proving that these algorithms have a bias towards KKT points of the corresponding margin maximization problem. We extend the analysis to Adam (without the stability constant), which maximizes the $\ell_\infty$ margin, and to Muon-Signum and Muon-Adam, which maximize a hybrid norm. Our experiments corroborate the theory and show that the identity of the margin maximized depends on the choice of optimizer. Overall, our results extend earlier lines of work on steepest descent in homogeneous models and momentum-based optimizers in linear models.