Implicit Bias

Current Vision-Language Models (VLMs) are celebrated for their multilingual capabilities, yet they operate under a flawed assumption: that one language corresponds to a single writing system. This overlooks billions of users of multi-script languages like Punjabi, Serbian, Hindi-Urdu, Kurdish, among many others, for whom a model's capability may be fractured by orthographic bias. We introduce PuMVR (Punjabi Multimodal Visual Reasoning), the first benchmark designed to quantify script-dependent bias through 375 culturally grounded image-reasoning tasks across Punjabi's three active scripts (Gurmukhi, Shahmukhi, Roman). Evaluating 10 state-of-the-art VLMs, we expose a substantial Script Gap: models frequently solve visual puzzles in one script while failing identical tasks in another, with accuracy deltas reaching 16% and Script Consistency Rates (SCR) as low as 24.8%. Crucially, visual input boosts absolute performance but does not close this gap, the relative bias persists. Our analysis suggests reasoning patterns show limited cross-script transferability, and Chain-of-Thought pathways diverge based on script alone. We propose SCR as a core metric for script-agnostic evaluation, challenging current multilingual assessment paradigms and providing a framework for equitable AI.

Generative artificial intelligence has the potential to improve productivity and transform the production of creative content. However, existing research indicates that image generation models are significantly influenced by biases. This work investigates the inherent biases and language-induced biases present in text-to-image models within the context of occupation-related image generation, complementing established metrics with human preference feedback. We present a comprehensive evaluation of five current text-to-image models: Midjourney v6.1, Stable Diffusion 3 Medium, DALL-E 3, Playground v2.5, and FLUX.1-dev , focusing on gender and ethnicity bias, image quality, and prompt alignment. To facilitate this evaluation, we developed the "Battle-Arena for Fair Image Synthesis" (BAFIS), a platform designed to collect human feedback on bias in generated images. Furthermore, we created a dataset comprising 21,140 synthetic images generated using multilingual prompts, which serves as a basis for our analysis. We further place our results within a broader social context by comparing them to official statistics from the German Federal Employment Agency. Our findings reveal systematic biases in text-to-image models, with established evaluation metrics in partial correlation with subjective user ratings. Thus, our research emphasizes the need for including human preferences to develop fairer and more inclusive text-to-image models.

Causal discovery methods commonly assume that all data is independently and identically distributed (i.i.d.) and that there are no unmeasured variables affecting the system. In practice, these assumptions are often violated, leading to inaccurate inference. In this paper, we study how to identify hidden confounding and selection biases from causal mechanism shifts. In particular, we show that structural biases lead to dependent mechanism shifts. That is, by considering for which variables the mechanisms change given data from different environments, we can tell which variables are unbiased, which are subject to hidden confounding, and which are undergoing selection bias. We formalize this into an empirically testable criterion based on mutual information, and show under which conditions it identifies structural biases. To tell which nodes are subject to what kind of bias, we introduce the StruBI algorithm. Experiments on synthetic and real-world data show that StruBI works well in practice, accurately recovering affected variable sets and types of biases, outperforming the state-of-the-art by a wide margin.

AI systems deployed in legal workflows hallucinate at rates that aggregate metrics report at ~52%, but this average conceals where errors concentrate and in which direction they run, leaving compliance officers without an actionable signal for trustworthy deployment. We present LegalHalluLens, an auditing framework with three components: typed hallucination profiles across four legally-motivated claim categories (numeric, temporal, obligation/entitlement, factual) over CUAD (Hendrycks et al., 2021); a Risk Direction Index (RDI) that reduces omission-versus-invention bias to a single deployment-comparable scalar; and a typed debate pipeline calibrated to both magnitudes and directions. Across 510 contracts and 249,252 clause-level instances we measure a within-model gap of approximately 38-40 pp between obligation/numeric and temporal claims that aggregate reporting hides, and show that two systems with matched 52% rates can carry opposite RDIs. The debate pipeline reduces fabricated detections by 45% with per-category gains tracking the diagnosis, matching commercial APIs with a substantially smaller backbone (4B active parameters). Typed profiles and RDI surface failure modes that aggregate metrics hide; we further show these diagnostics serve as calibration inputs for multi-agent debate pipelines, where Skeptic challenges and asymmetric gates targeted at measured failure modes outperform generically-tuned debate. The framework supports direction-aware procurement, accountability, and agent design for legal AI deployed in the wild.

Depression is the leading cause of disability worldwide, and early detection of symptom change is essential for timely intervention. Validated instruments such as the Patient Health Questionnaire-9 (PHQ-9) support symptom monitoring at scale, but real-world completion rates are low, introducing response bias and systematic missingness. Passive approaches that infer severity from routinely generated data could close this gap. We address this by predicting PHQ-9 total scores directly from transcripts of conversations between users and an AI mental health application, requiring only conversation text and no additional clinical data. We fine-tune a Qwen3.5-27B backbone with a regression head, augment 3,111 ground-truth labels with pseudolabels generated by a reasoning model (Claude Opus) and iteratively trained intermediate models, for a combined dataset of 6,283 users. On a held-out test set of 842 users, our best model achieves MAE = 2.6, RMSE = 4.0, Pearson r = 0.80, and AUC = 0.91 at the PHQ-9 >= 10 clinical threshold. We also find AUC > 0.87 at every severity threshold from PHQ-9 >= 3 to PHQ-9 >= 24, demonstrating that the model captures depression severity across the full clinical spectrum. This work opens the door to passive, continuous symptom monitoring in AI mental health platforms, without requiring users to complete self-report measures.

Evaluations of social bias in LLMs largely focus on whether models generate or imply biased content. However, as LLMs are increasingly used as judges of bias, they may exhibit social biases in subtler ways in how they evaluate biased content, which current methods do not systematically capture. We call this second-order bias: social bias in an LLM's judgment about social bias, which we evaluate through a novel, philosophically grounded reasoning task. Drawing on entitlement epistemology, we conceptualize bias as misplaced foundational knowledge that shapes an agent's rational inquiry, and derive a logical reasoning task for LLMs to judge to whom a biased text is acceptable or non-acceptable. We develop two simple metrics to measure how biased LLM judges are in inferring demographics for acceptability without sufficient support, and how these inferences vary across groups targeted by biased texts. Evaluating open and closed models, we find that our task evades safety guardrails by surfacing bias in model judgment. It varies systematically across target groups, reflects implicit social maps, and shows how models are still triggered by demographic labels. Our work points to the need for LLM bias evaluation in judgment tasks and broadly, for more theoretically grounded approaches to bias evaluation in NLP. We release our code and model responses at https://github.com/uofthcdslab/second-order-bias.

Large language models (LLMs) are becoming a major way for consumers to find products, but we do not yet understand how brands compete in this new channel. We study brand dynamics in LLM recommendations using skincare products -- a category where consumers cannot easily judge quality before buying and must rely on brand reputation -- across three commercial LLMs (GPT-4o-mini, Claude Sonnet, Gemini 3 Flash), with a robustness check on search goods. In three experiments, we find: (1) a Conditional Monopoly where well-known brands get recommended 100% of the time (IAI = 10.0) when all products have the same specifications, but this dominance disappears with less than a +0.1-star rating advantage for a competitor; (2) authority-style marketing language, including fabricated clinical-evidence claims, breaks this monopoly at a Bias Surplus Value equal to +0.17 rating points, with each model responding differently; and (3) a social dilemma in multi-brand GEO competition: when all brands adopt the same optimization strategy, individual payoff falls from +0.802 to +0.007 in our payoff proxy, and non-participating brands receive zero recommendations in our tests. Our results suggest that generative engine optimization (GEO) should be studied not only as a security risk, but also as an emerging marketing practice that shapes market competition.

As Large Language Models (LLMs) become embedded in everyday communication, capturing regional linguistic variation is essential for reliable and equitable language use. In Portuguese, European (pt-PT) and Brazilian (pt-BR) varieties remain unevenly represented, with pt-BR dominating in data quantity, while LLM preference for Portuguese variants remains underexplored. To address this gap, we introduce P3B3, an expert-curated language variety agnostic benchmark of conversational prompts, along with an evaluation framework for measuring variety bias and controllability. Experiments on several models show that most LLMs exhibit a strong bias toward pt-BR, with variation in controllability across models. These results highlight the need for more balanced multilingual representation across language varieties.

Knowledge-based visual question answering (KB-VQA) lets vision-language systems answer questions that exceed their parametric knowledge by conditioning a reader on passages retrieved from a Wikipedia-derived knowledge base. In pure-text long-context LLMs, retrieved-context use follows the U-shaped lost-in-the-middle effect of Liu et al. (2024): information at the start and end of context is used, the middle is lost. Whether this transfers to deployed multimodal KB-VQA is open. To close this gap, we design the first controlled probe of reader-side position dependence in multimodal KB-VQA: a gold-position protocol in which only the gold passage's prompt slot varies within question. We run it on three open-source 7B/8B VLM readers and two KB-VQA benchmarks with up to 20 retrieved passages. The shape flips from U to primacy: gold-at-first beats gold-at-last by 16 to 26 points on all six combinations of reader and benchmark, an effect we call Lost at the End; the gap holds at every scale we test, 3B to 32B, attenuating at 32B. Three targeted ablations narrow the cause. A text-only control that removes the image and changes nothing else shows the primacy is already present in text mode and does not depend on the image. Image-position and distractor-shuffle ablations trace the effect to prompt slot 0 of the instruction-tuned reader, where a second answer-bearing passage placed later is largely wasted. On a frozen reader, three retrieval-side fixes (MMR, oracle reranking, rank-based reordering) all fail to improve on the deployment default. Our findings indicate that recall@k is the wrong metric for deployed KB-VQA and that the remaining headroom sits on the reader side; we release our protocol as a controlled instrument for evaluating reader-side interventions.

Reasoning models are increasingly used in settings where the final answer is not the only object of review: educational tools may show students intermediate steps, decision-support systems may require human oversight, and audit workflows may inspect traces for misleading or biased input. In such settings, two responses can receive the same final-answer score while differing in whether the trace explicitly flags injected biasing content. Accuracy-only evaluation collapses these cases. We study this gap as a measurement blind spot for responsible evaluation and introduce a minimal trace-level diagnostic with two axes: \emph{susceptibility} (whether the bias breaks a previously correct answer) and \emph{acknowledgment} (whether the trace contains a rubric-defined surface reference to the injected content). Across thousands of biased GSM8K trials, GPT-4o and Claude Sonnet~4 have similar susceptibility rates ($1.3\%$ vs. $1.2\%$) but substantially different acknowledgment rates ($13.0\%$ vs. $75.0\%$) under the same rubric.

Drug-target interaction (DTI) and affinity (DTA) predictors increasingly achieve strong benchmark scores, yet their internal use of sequence, fingerprint, and graph features often remains opaque. We present an interpretability audit of BridgeDPI architecture on three different datasets including Gao, Human, and C.elegans. This study combines gradient-based attributions -- integrated gradients, saliency, layer-wise relevance propagation, SmoothGrad, and SmoothGrad-IG -- with feature-wise occlusion ablation and strict intersection consensus across methods to reduce single-explainer bias. We summarize sensitivity and signed effects at raw inputs, at the bridge similarity scaffold, and through the graph convolution, including edge-level sensitivities and targeted edge removals. The results show that explainability is most informative when treated as model criticism: it reveals modality dominance, padding and special-token artifacts, dataset-dependent cooperative versus suppressive effects across layers, and chemistry-consistent fragment and composition motifs where methods agree. These analyses do not substitute for structural or experimental ground truth, yet they can provide testable hypotheses for downstream validation in computational drug discovery pipelines. More broadly, applying modern XAI to contemporary DTI/DTA models is still an early pass over the rich structure implicit in trained weights and data -- yet even this first layer of scrutiny already helps researchers relate predictions to drug- and target-side representations and to prioritize external validation.

Deployed knowledge-tracing models are typically frozen after training, yet systematic per-item logit bias arises, from limited per-item expressivity in backbone architectures and from post-deployment shifts in item properties, degrading prediction quality. Global post-hoc calibrators such as Platt scaling, temperature scaling, and isotonic regression improve probability estimates but leave discriminative ability, as measured by AUC, unchanged. This AUC invariance is a structural consequence of monotone score-only transforms; recovering the stranded discrimination requires conditioning on item identity. We propose SLC (State-space Logit Correction), which converts binary observations to Gaussian pseudo-observations via Laplace/IRLS, applies empirical-Bayes shrinkage through a Kalman smoother, and fits an offset-Platt link. The state-space formulation also yields a detectability bound that characterizes the Bernoulli information floor, explaining why temporal tracking provides no benefit at current data densities. Across four datasets, five backbones, and three seeds, SLC improves AUC on all four datasets and NLL on three, with the advantage concentrating on sparse items. Cross-domain controls suggest that the same phenomenon can arise beyond education when the deployed backbone leaves entity-level bias.

Editing pretrained neural networks requires specialized algorithms tailored to specific objectives. Designing such algorithms is often time-consuming and demands significant effort. We present an exploratory framework that formulates neural model editing as a reinforcement learning problem, where agents modify models using reward feedback. We introduce two environments: MaskWorld, where agents scale weights multiplicatively, and ShiftWorld, where agents apply additive weight updates. The reward function combines a utility-preservation objective with a task-specific editing objective, enabling agents to learn targeted modifications while maintaining overall model performance. We evaluate the framework on bias mitigation in text classification and machine unlearning in image classification, both of which traditionally rely on specialized algorithms. Our results show that the learned policies reduce forget set accuracy to nearly 0% while preserving over 90% retain set accuracy on the unlearning task. In the bias mitigation setting, the learned policies improve bias-related performance by more than 5% while maintaining general classification utility. Our findings show that neural model editing can be cast as a reinforcement learning problem, allowing editing policies to be learned from reward feedback rather than manually engineered for each task.

Learning grounded word meaning from natural experience requires resolving two ambiguities in infant-view recordings: when the named referent appears and where it is in a cluttered frame. In SAYCam-style data, caregiver speech is sparse and weakly synchronized with egocentric video, so single-frame contrastive pairing yields noisy positives in which the intended object is absent or entangled with distractors. We propose BabyMind, an object-first bias for child-view contrastive learning under sparse, noisy supervision. BabyMind extracts candidate object embeddings using an offline mask-based region interface, links candidates across a short utterance-centered window into lightweight object files via tracking, and aligns utterances to bags of object files with a prototype-space multiple-instance contrastive objective. Track-coherence and global-object agreement regularizers stabilize learning and transfer object-file structure into the global frame embedding used at evaluation. On SAYCam-S, BabyMind improves Labeled-S 15 forced-choice accuracy by +2.6 points over CVCL and yields consistent gains on in-vocabulary out-of-distribution benchmarks. Code is available at https://github.com/sathiiii/BabyMind.

Anticipating LLM behavioral tendencies from low-cost psychometric probes is critical for safe deployment, but only if self-reports (SR) reliably predict behavior. Recent work documented substantial SR-behavior dissociation in LLMs, but relied on broad personality traits (Big 5) that predict specific behaviors weakly, even in humans. Furthermore, the isolation of conversational sessions combined with weak context matching left open whether LLMs truly lack coherence or whether the conditions needed to detect such coherence were not met. We contrast Big 5 with the Theory of Planned Behavior (TPB), which measures intention targeted to a specific behavior and predicts human behavior substantially better than broad traits. We run experiments across four behavioral tasks and 11 frontier LLMs, while also varying session context and identity induction. We find that SR-behavior coherence exists but is selective. 1) Within a shared conversation, the Theory of Planned Behavior reaches human-level coherence; Big 5 does not. 2) Across separate conversations, coherence survives only for behaviors anchored outside the immediate prompt, such as implicit bias shaped by training, and collapses when behavior is strongly primed by context, as with sycophancy. 3) Persona prompting makes self-reports more consistent across conversations, but does not bring behavior into alignment. These findings suggest that coarse personality frameworks, such as Big 5 may not be the best tools for testing deployment behavior. More task- and behavior-specific instruments are needed, and even these must be evaluated across tasks and contexts.

As automatic speech recognition (ASR) systems shift toward multilingual support and low-resource language modeling, phoneme-based layers serve as a critical language-agnostic foundation. However, most evaluations of ASR's demographic biases related to race, age, gender, and accent focus on standard grapheme-based ASR systems with comparatively little emphasis on phoneme-based systems. In this study, we evaluate the performance of WhisperIPA and ZIPA, two state-of-the-art open-source systems that generate International Phonetic Alphabet (IPA) transcriptions. Our evaluation includes existing multilingual speech corpora and demographically annotated English-language corpora, comparing model-generated IPA transcriptions against grapheme-to-phoneme (G2P) systems using both standard phoneme error rate (PER) and a proposed Soft PER metric that tolerates linguistically similar phoneme substitutions. Our analysis examines how performance varies across language, gender, accent, ethnicity, and age, revealing persistent disparities even after accounting for acceptable phonemic variation. These findings, while limited, provide insight into potential sources of bias and inform the development of more inclusive and linguistically robust phoneme-based ASR systems. Our code and data are publicly available.

Warning: This paper contains several toxic and offensive statements. Modern large language models (LLMs) are typically aligned through large-scale post-training to ensure fair and reliable behavior. In this work, we investigate how easily such guardrails can be broken by Group Relative Policy Optimization (GRPO). We show that one-shot GRPO training on a single biased example is sufficient to induce systematic bias, with stereotype-driven reasoning generalizing across attributes, categories, and benchmarks. We further find that models differ in their susceptibility based on the initial likelihood of producing biased outputs. Our results reveal a critical vulnerability in post-training: alignment can be overridden by a single example.

Transformer-based models have emerged as leading paradigms in time-series forecasting in recent years, employing self-attention mechanisms to capture long-range dependencies. Despite their success, these single-stage forecasting architectures exhibit persistent systematic residual biases arising from structural discrepancies, unmodeled stochastic components, or inadequate multi-scale temporal representations. This limitation persists when residuals are treated as irreducible noise, precluding adaptive correction of structured error patterns. To address this limitation, we introduce a two-stage, model-agnostic framework that explicitly decouples forecasting and residual learning into distinct stages of representation learning. A base transformer first generates the initial predictions. Subsequently, a dedicated meta-corrector dynamically models structured error patterns across multivariate channels, preserves cross-variable dependencies, and iteratively refines the residual bias of the base transformer. By formalizing this pipeline as a hypothesis space expansion, our framework addresses approximation limitations inherent in single-stage architectures, removes reliance on restrictive assumptions, and enables end-to-end learning of complex error dynamics. Evaluated on eight popular benchmark datasets using established protocols, our approach achieves state-of-the-art performance, with significant improvements in standard metrics (MSE, MAE). The results demonstrate the framework's ability to mitigate systematic biases and enhance robustness to complex temporal dynamics, advancing the practical applicability of transformer-based forecasting models.

Pairwise comparisons combined with aggregation methods like Elo have become central to evaluating generative models, yet concerns remain that they reward superficial stylistic cues or display judge biases. In a more positive turn, we show that model rankings from pairwise comparisons strongly agree with ground-truth-based accuracy rankings when such ground truth is available for comparison. By converting five well-known benchmarks into free-form generative evaluations, we find that Elo rankings achieve a Spearman correlation above 0.9 with accuracy rankings and substantially outperform direct evaluation when the judge is weak. Furthermore, style and judge bias have only minor effects on model rankings, despite most judgments occurring on pairs where both candidate answers are correct (or incorrect). On such pairs, we find that repetition after the final answer (echo) is a causal driver of judge preference.

Federated fine-tuning of foundation models using Low-Rank Adaptation (LoRA) offers a communication efficient solution for distributed learning. However, existing federated LoRA methods suffer from two fundamental limitations: (1) structural aggregation bias, where independently averaging low rank factors fails to approximate the true combined update, and (2) client side initialization lag, as clients repeatedly reinitialize LoRA parameters across communication rounds, slowing convergence. We propose HyperLoRA, a unified framework that addresses both issues through amortized federated adaptation through hypernetwork-driven LoRA generation and product space aggregation. Instead of iterative per-client optimization, HyperLoRA employs a learned generator that maps client distribution signatures to LoRA initializations, effectively amortizing per client adaptation. On the server side, we introduce a learned aggregation module that directly synthesizes updates in the low-rank product space, eliminating the inconsistencies of factor-wise averaging. A lightweight residual correction module further improves stability under heterogenous (non-IID) client distributions.By replacing iterative optimization and heuristic averaging with learned operators, HyperLoRA jointly enables efficient personalization, unbiased aggregation, and faster convergence. Experiments on federated vision and vision-language benchmarks show that HyperLoRA achieves improved convergence speed, greater robustness to distribution shift, and stronger personalization performance compared to prior federated LoRA methods.

Current evaluations for Multimodal Large Language Models (MLLMs) overwhelmingly focus on utility-driven objectives, leaving model behavior under logic-neutral scenarios largely underexplored. Stochasticity is essential in scenarios where multiple actions are equally valid, such as recommending travel itineraries or daily schedules where multiple options have similar utility. In such settings, deterministic policies may lead to repetitive behaviors and reduced coverage of valid alternatives. To bridge this gap, we propose RandomBench, a benchmark designed to evaluate whether MLLMs can maintain distributionally neutral behavior when selecting among equivalent options. We further introduce three metrics, including RI, BCI, BII, to quantify entropy and distributional bias. Experiments reveal a pervasive phenomenon termed Stochastic Collapse, where MLLMs fail to maintain uniform randomness under explicit random instructions, with top-1 probabilities reaching 97% from the ideal one quarter baseline and RI dropping to 0.068 in Claude Sonnet 4.6. Extensive ablation studies further demonstrate that these deviations persist across languages and representation formats, highlighting the robustness of distributional collapse in logic-neutral decision settings.

We investigate the extent to which the language processing of LLMs resembles human cognitive processes, focusing on a human cognitive bias called the $\textit{neglect-zero effect}$. This effect refers to the human tendency to ignore $\textit{zero-models}$, which are configurations that render a proposition vacuously true by virtue of an empty set. We focus on two types of inferences driven by the neglect-zero effect, and examine how LLMs process these inferences by comparing their behavior with that in an inference that does not involve the neglect-zero effect. For this purpose, we employ a paradigm based on $\textit{structural priming}$, where recent exposure to a preceding sentence (the $\textit{prime}$) facilitates the processing of a subsequent sentence (the $\textit{target}$) due to their structural similarity. We prepare primes to force LLMs to consider the zero-model, and analyze whether they also consider it in the target. The results suggest that the neglect-zero effect may not occur in the LLMs analyzed in this study. Our code is available at https://github.com/ynklab/neglect_zero

Rubric-based reinforcement learning (RL) uses an LLM-as-a-Judge (LaaJ) to score model outputs according to rubrics as rewards. However, policy models may exploit latent biases in the judge, leading to reward hacking and ineffective or unsafe training outcomes. In real-world rubric-based RL, such hacking behaviors are often subtle and entangled with multiple judge biases, making them difficult to analyze, detect, and mitigate. In this paper, we introduce CHERRL, a controllable hacking environment for rubric-based RL. By injecting known biases into LaaJ, CHERRL enables stable reproduction of reward hacking, explicit observation of reward divergence, and precise identification of hacking onset. This provides a clean experimental testbed for studying the mechanisms and mitigations of reward hacking in rubric-based RL. To demonstrate its utility, we analyze different judge biases from the perspectives of discoverability and exploitability, and explore an agent-based system for automatically detecting reward hacking onset from training logs. The code and environment are publicly available at https://github.com/THUAIS-Lab/CHERRL.

CutMix has become the de facto standard mixing augmentation, yet its label assignment rests on a flawed assumption: The area of the pasted patch faithfully reflects its semantic contribution to the mixed image. In practice, however, patches frequently land on background regions, assigning label credit to classes whose objects are not visible. The mean discrepancy of the CutMix label and the semantic object area is $21.5\%$. In $17\%$ of samples an image contributes zero visible object pixels yet receives nonzero label weight. We propose Object-Aware CutMix (OA-CutMix), which corrects this bias by replacing the area-based CutMix weight with one derived from precomputed segmentation masks, assigning labels in proportion to the visible object area each image contributes to the mix. The image mixing procedure is left entirely unchanged. We evaluate OA-CutMix against 10+ static and dynamic mixing methods across 4 architectures and 6 datasets. OA-CutMix consistently achieves the highest accuracy over all tasks, outperforming even dynamic mixing methods, but at a fraction of the training-time cost. Improvements are largest for small objects, where the label bias from CutMix is greatest. Thus, correcting the label is sufficient to match or exceed the performance of methods modifying the image mixing algorithm.

Multimodal Large Language Models (MLLMs) are increasingly used for video understanding, yet their reliability under multi-video inputs remains poorly understood. We study positional bias in multi-video summarization, where the quality of a per-video summary can change with the video's input slot even when the underlying content is unchanged. We construct a benchmark from ActivityNet and News videos, covering Cooking, Domestic, Leisure, and News settings with two- and four-video inputs. We evaluate nine open-source and proprietary MLLMs and measure position effects with three complementary metrics: Coverage, Directional Positional Bias (DPB), and Middle-Edge Gap (MEG). Our results show that positional effects are domain- and model-dependent: signed directional bias can be small even when middle positions underperform, and increasing visual or generation budget does not uniformly remove the imbalance. We further analyze prompt-level mitigation methods. Together, the results show that multi-video summarization remains sensitive to input protocol and position, motivating more robust order-invariant multimodal systems.

In this study, we evaluate the performance of skin lesion classification using ResNet-based convolutional models, focusing on the impact of demographic bias in training data, particularly variations in patient sex and age. We use linear programming to generate datasets with controlled demographic characteristics, allowing systematic investigation of bias effects. Three learning strategies are evaluated: a single-task model, a reinforcing multi-task model, and an adversarial learning scheme. Our sex-based analysis indicates that sex-specific training datasets optimise model performance. Notably, including male patients in the training data improved performance for the male subgroup, even in female-majority cases. Reinforcing and adversarial learning schemes narrowed or eliminated bias gaps in balanced and female-majority datasets. However, these strategies proved less effective in male-majority settings, where models continued to perform better for males than females. The two learning schemes showed marginal bias reduction compared to the baseline model in predominantly male patient populations. Age-based analysis demonstrates comparable baseline performance across the three model approaches, with performance declining across age categories. Younger groups consistently achieve the highest performance, regardless of training data distribution. Although balanced training yields optimal results for the youngest age category, performance decreases in older categories. We find that sex biases arise mainly from data imbalances, while age biases consistently favour younger groups regardless of distribution. These distinct mechanisms require targeted mitigation strategies. Additionally, cross-dataset validation on two external datasets revealed that domain shifts notably affect performance and patterns of demographic bias.

Dense retrieval compresses a passage into a single vector, but this compression is positionally skewed: early content dominates the embedding, and retrieval degrades when the relevant span appears later. Prior work proposed an inference-time method that counteracts this skew by equalizing the pooling token's attention across passage segments. However, (i) it redistributes attention at a fixed strength, (ii) it forces the pooling token's attention to itself to a fixed basket-level mass despite substantial variation across layers and architectures, and (iii) its effect on retrieval has not been evaluated. We introduce a strength coefficient that interpolates between uncalibrated and fully equalized attention, together with an efficient implementation that reduces peak calibration memory overhead from 5-7 GiB to under 1 MiB. Across three embedding models and two pooling schemes, moderate calibration provides a better retrieval trade-off than full equalization. We introduce a variant that preserves the pooling token's self-attention mass and redistributes only the remaining mass. On a position-aware retrieval benchmark spanning 10 languages and 31 domains, a configuration selected on English FineWeb-PosQ and transferred without tuning reduces position sensitivity in all 16 evaluated length-quartile, model, and retrieval-setting combinations, by up to 43% relative, while improving nDCG@10 by up to 4.8% relative and leaving general retrieval effectiveness on NanoBEIR essentially unchanged. Calibration runs at indexing time, adding no query-time latency. We release our code at github.com/impresso/fair-sentence-transformers

Entropy minimization (EM) is the dominant objective for test-time adaptation, yet its failure mode, model collapse, remains poorly understood. In this work, we show that distribution shifts can cause feature clusters corresponding to distinct classes in the model's representation space to merge, while the decision boundary remains fixed. This induces a systematic skew in the predicted class distribution, referred to as prediction bias. Prediction bias refers to a shift in the predicted class distribution, with some classes overrepresented and others suppressed. We show that entropy minimization amplifies this prediction bias by tightening the existing clusters, reinforcing the incorrect groupings until all predictions collapse to a trivial solution. Next, to demonstrate the significance of prediction bias and mitigate it, we further propose Distribution Shift Bias Reduction (DSBR), a bias-correcting objective that specifically targets this failure mode by equalizing the contribution of each predicted class to the unsupervised entropy minimization loss. To study this failure mode, we design suitable adaptation settings using four medical-imaging datasets and additionally evaluate on ImageNet-C. We find that DSBR consistently stabilizes test-time adaptation, prevents model collapse, and matches or outperforms state-of-the-art methods. Moreover, DSBR operates solely at test-time.

Massive activation spikes in Large Language Models (LLMs) severely degrade quantization by stretching dynamic ranges. While prior hypotheses characterize these as high-level scalar biases, we argue that they are merely the scalar intermediates of rigid, structural vector biases in the spike-carrying tokens. We show that these tokens converge to constant vectors after normalization that drive the attention sink and value-state drain mechanisms. We geometrically substantiate this by analyzing the coordination of projection weights: $W_K$ contrastively amplifies the vector, $W_Q$ aligns semantic tokens toward it, and $W_V$ projects it into the spectral null-space. Furthermore, we reveal that the model actively preserves these structural biases against Rotary Positional Embedding (RoPE) perturbations by localizing them in "zones of rotational stability" utilizing low-frequency bands and coherent channel pairs. Leveraging this, we propose INSERTQUANT, a post-training quantization (PTQ) framework that clamps spikes and restores their function via pre-computed template vectors. This renders activations strictly spike-free, enabling robust low-bit quantization with high fidelity. INSERTQUANT achieves parity with state-of-the-art per-tensor quantization methods on LLMs and uniquely generalizes beyond text to other modalities such as ViTs.

Generative emulators of protein dynamics produce plausible trajectories at a fraction of the cost of molecular dynamics, but they inherit their training distribution and tend to revisit known states rather than reach rare ones under long-horizon extrapolation. Inspired by classical enhanced sampling, we introduce an implicit, history-dependent bias in the generative space of a pretrained emulator. Specifically, a history-aware score estimator augments the frozen emulator with a distance-weighted bias that steers reverse-time sampling away from previously generated structures, regularized by an environment-support term. To preserve structural validity at long horizons, a score-based refinement step re-projects drifted samples onto the data manifold using the frozen emulator. Our experiments demonstrate that the method (i) raises diversity by $35\%$ on DynamicPDB-80; (ii) on $12$ zero-shot Fast-Folding proteins, the learned bias alone reaches the unbiased emulator's coverage up to ${\sim}15\times$ faster, and pairing it with refinement reaches the coverage up to ${\sim}37\times$ faster while covering ${\sim}3\times$ as many low-energy states. Code will be released soon.

Text-Based Person Search (TBPS) aims to retrieve pedestrian images using natural language queries. However, existing TBPS models, especially those based on CLIP, struggle with fine-grained understanding due to global representational bias and semantic sparsity inherited from training on short captions. This results in weak fine-grained alignment, exacerbated by the scarcity of region-level annotations. To address this, we propose ROGLE (Robust Global-Local Embedding), a unified framework that overcomes reliance on costly manual annotations through an automated Region-to-Sentence Matching (RSM) strategy. RSM automatically mines pseudo region-sentence pairs for scalable fine-grained supervision. Furthermore, ROGLE employs a multi-granular learning strategy that fuses global contrastive learning with region-level local alignment. We also introduce the P-VLG Benchmark, a large-scale dataset constructed by curating and enriching images from established public benchmarks. It features over 100,000 annotated regions and rich long-form captions, making it the first TBPS benchmark to support both global and local assessment protocols. Extensive experiments show that ROGLE significantly outperforms existing approaches, particularly on challenging long-form queries. Code and the P-VLG benchmark will be made publicly available.

Conversational tutoring agents have been shown to improve learning engagement and student outcomes, and large language models (LLMs) are increasingly used in these systems to provide scalable, personalized feedback. However, LLMs may perpetuate or amplify stereotypical social biases, posing particular risks in educational settings. In this study, we evaluate LLMs in conversational tutoring scenarios to identify high-confidence social biases, instances where models are unable to identify biased judgments in tutoring conversations while maintaining strong confidence in their assessments, potentially affecting their reasoning and the feedback they provide to learners. We present a new dataset generation method that enables bias evaluation under naturalistic instructional conditions by regenerating student-AI tutor interactions and introducing turns with controlled bias derived from a benchmark dataset. Using this data, we assess multiple LLMs' ability to detect stereotypical biases and analyze the confidence and reasoning underlying their responses through computational and human evaluations. We find that bias detection is substantially more challenging in conversational tutoring contexts than in benchmark-based evaluations, and that state-of-the-art LLMs are overconfident in their incorrect assessments of stereotypical bias statements. Moreover, model confidence strongly influences reasoning and feedback, highlighting the risks of overconfident, biased behavior in LLM-based tutoring agents. We conclude by discussing implications, mitigation considerations, and directions for future research.

Studies of human reasoning have shown that people are typically stronger at evaluating reasoning than producing it from scratch. In contrast, large reasoning models (LRMs) are trained to excel at producing long chains of reasoning to solve complex problems. How then do LRMs perform at evaluating reasons? We investigate this with the Valid-Answer-Invalid-Reasoning (VAIR) dataset: math problems and solutions with trivial reasoning flaws but valid answers, designed to isolate reasoning evaluation from the confound of reasoning production. Unlike humans, who we find are only 6% worse at grading than solving such problems, we find a substantial production-evaluation gap in LRMs: frontier models score as low as 48% when evaluating VAIR solutions, despite near-perfect solution production. Why this enigma? Through chain-of-thought (CoT) analysis, we find evidence of an answer confirmation bias: LRMs often produce then check for the correct answer instead of carefully verifying each step, fabricating rationalizations even when noticing anomalous reasoning. Linear probes corroborate this, showing that while LRM activations encode some representation of valid reasoning, they fail to robustly represent VAIR solutions as invalid. Causal patching of the final answer's representations causes LRM verdicts and activations to flip, demonstrating that answer validity is responsible for models' confirmation biases. These findings indicate an outstanding limitation in dominant approaches to reasoning training, which incentivize LRMs to produce and confirm reasoning towards correct answers, but not to robustly evaluate the underlying reasons.

Transit network design depends not only on the optimization algorithm but also on who shows up to the public hearing. Current practice often collects one-directional comments from self-selected attendees, leaving participant mix as an uncontrolled source of outcome variation. We present AGORA, a framework that holds the network, demand, and solver fixed while systematically varying meeting composition through stakeholder agents, structured deliberation, and governance gates. Across two standard benchmark networks at different scales, we find that (i) aggregate outcomes vary little across compositions, but on tail risk and fairness disparity, representative sampling still tends to outperform skewed compositions; (ii) without deliberation, composition produces no variation at all, showing that deliberation is the mechanism through which who attends affects outcomes; and (iii) governance gates compress cross-profile variance without shifting the average outcome on Mandl, but low acceptance on Mumford0 shows thresholds require instance-specific calibration. These findings reframe participation bias from an uncontrollable input to a process-design problem: even without guaranteed representative attendance, well-structured deliberation and governance criteria can substantially reduce how much outcomes depend on who is in the room.

This letter studies CSI denoising for MIMO--OFDM with variable NR resource block (RB) allocations. ReFLEX is a length-generalizable Transformer whose frequency attention uses a relative-frequency position bias (RFPB) generated from subcarrier offsets. A single checkpoint handles unseen RB lengths and can be applied to sparse DM-RS observations in the tested RB5/RB10 PUSCH setup without retraining. In a 3GPPTR38.901 UMa NLOS channel, ReFLEX achieves about $-9.6$dB NMSE on unseen RB lengths. In NR PUSCH/UL-SCH simulations, ReFLEX denoising followed by time-frequency interpolation reduces the 10% BLER threshold by about 2--3dB.

Given the surge of harmful AI-generated imagery online, reliably distinguishing authentic images from generated ones has become an urgent research topic. While many proposed detection methods perform well under controlled settings, they often collapse when tested on real-world data. A potential root cause are subtle biases in the detectors' training data. As a result, detectors may rely on spurious correlations instead of learning true forensic artifacts. While a recent line of work has identified the problem, there is not yet an established protocol to evaluate how biased a detector actually is. In this work, we therefore take a step back: First, we discuss what it means for a detector to be biased, and how this differs from a lack of robustness. Second, we propose BIAS-ID, a transparent framework for analyzing and quantifying the presence of transformation biases in AI-generated image detectors. We validate our framework by performing an evaluation of six detectors across two datasets, revealing that several state-of-the-art detection methods are strongly affected by biases. Our results highlight the importance of bias-aware evaluation for developing reliable AI-generated image detectors.

Large language models (LLMs) can reveal and amplify societal biases during chain-of-thought (CoT) generation. We present COFT (Chain of Fair Thought), a training-free decoding method that applies token-level fairness control at decode time, with distribution-free marginal validity guarantees (under exchangeability) for any frozen causal language model. COFT operates in three stages. First, it creates a masked counterfactual prompt by replacing sensitive spans with neutral tokens. Second, it compares the factual and masked logit distributions through lightweight logit fusion to attenuate attribute-driven biases. Third, it uses dual-branch split-conformal calibration to certify per-step candidate token sets at a user-chosen risk level. We evaluate COFT across six models and multiple bias benchmarks. Our method reduces standard bias metrics by 30-55% (median 38%) while preserving task utility and language quality. Reasoning accuracies remain unchanged within run-to-run noise margins. The computational overhead is modest, equivalent to one additional cached forward pass (<=11%). COFT offers a clear, auditable path to safer CoT generation with significant bias reduction, negligible utility loss, and no requirement for retraining, auxiliary classifiers, or weight access.

As AI-generated and AI-assisted content floods online spaces, source labels attached to such content can distort human reasoning judgments, with downstream consequences for moderation, evaluation, and decision-making. Whether LLMs share this vulnerability, or offer more source-agnostic evaluation, remains an open question with direct implications for human-AI collaboration. We examine this issue using logical fallacies as a controlled setting to isolate source-label effects on reasoning quality, independent of domain knowledge. We conduct an online study (N=505) where participants are assigned to a source condition (human, AI, human with AI assistance, AI with human assistance, or no disclosure) and evaluate comments containing logical fallacies, comparing their judgments with those of LLMs (GPT-5.2, Gemini 2.5 Flash, Claude Sonnet 4.5), who were evaluated across the same source conditions. Human evaluators were significantly more susceptible to fallacies labeled as written by human or human with AI assistance and assigned higher trust and evaluation ratings in these conditions. LLM evaluations remained comparatively stable across source labels, though performance varied across models. Confidence levels were similarly high across conditions for both humans and LLMs, regardless of fallacy presence. Our findings indicate that source-label bias in reasoning evaluation is primarily a human vulnerability and highlight the potential of human-LLM collaboration in increasingly AI-mediated environments.

Self-supervised learning (SSL) has produced a diverse landscape of vision transformers (ViTs) whose pretrained representations support a wide range of downstream tasks. Towards a better understanding of these models, a body of work has assessed the mechanics of their self-attention as well as the types of information captured across their representations, revealing, for example, stark differences between models trained with contrastive learning (CL) and masked image modeling (MIM). However, the total of these advances on model understanding has to date not yet fully permeated a larger community, where, e.g., insights that are specific to CL models are still at times generalized to MIM models. To make model understanding straightforward and intuitive for a broad community, we propose a simple and easily interpretable visualization protocol. Our protocol is based on visualizing unsupervised semantic segmentation results, yet by no means do we focus on top segmentation performance. Instead, our protocol allows us to easily convey model behavior that consistently emerges across images. Benchmarked on a diverse set of SSL models across layers and representations, our protocol allows us to gain novel insights into distinct positional biases and scaling behaviors, including, e.g., strong boundary artifacts in DINOv3-Large model tokens. These novel insights come on top of more easily conveying a range of previous findings. Our protocol further allows us to clearly visually convey and distinguish between positional effects and the closely related but distinct locality bias, the latter being much more extensively studied in the literature so far. Our protocol is publicly available, serving to catalyze further model understanding for a broad community.

Vision classifiers can exploit spurious correlations, achieving high in-distribution accuracy yet failing under distribution shift. Existing approaches to bias mitigation and analysis often depend on curated datasets, spurious-attribute or group labels, or retraining, which may be infeasible once a model is deployed or the relevant bias is unknown. We present a bias-label-free, post-hoc method for identifying spurious concepts in frozen vision models, relying only on standard class labels from a held-out audit dataset. For each target class, we collect patches from inputs predicted as that class and apply non-negative matrix factorization to intermediate activations to obtain a bank of interpretable concept vectors. Candidate concepts are then ranked with a bias estimator derived from their interaction with backpropagated gradients on misclassified examples: bias concepts tend to get activated when correcting false negatives and suppressed when correcting false positives. On Colored MNIST and Waterbirds the method recovers concepts aligned with the known spurious cue, and on CelebA it surfaces decision-relevant directions that only partially coincide with the annotated gender attribute; suppressing the top-ranked concepts at inference time improves worst-group accuracy by up to 17.9 percentage points on Waterbirds and 10.4 on CelebA without any retraining or parameter updates. Our method identifies decision-relevant spurious directions that need not coincide with annotated ones, providing both an interpretable auditing tool and an actionable debiasing handle for frozen vision models. Code is available at https://github.com/vitryt/label-free-bias-identification.

LLM-powered social agents are increasingly used to simulate online social behavior, yet their realism remains difficult to validate. Existing work has largely relied on general-purpose benchmarks, while less attention has been paid to short, reactive discourse such as audience replies to online news. In this paper, we evaluate whether LLM-generated reactions to Spanish online news reproduce measurable properties of real audience discourse. Using the Hatemedia dataset, we pair 5,631 news items with 58,555 real audience reactions, and generate a matched synthetic dataset using five LLMs under a shared experimental setting. We compare real and synthetic reactions across three dimensions: hate speech, sentiment, and semantic alignment, considering both off-the-shelf and fine-tuned generation. Results show that off-the-shelf models are poor proxies for real audience reactions: they strongly underproduce hate speech, introduce model-specific sentiment biases, and remain distributionally distant from human replies. Fine-tuning improves fidelity unevenly. Qwen3 provides the most balanced approximation, while Mistral7B achieves the strongest sentiment and semantic alignment but overshoots hate prevalence. Plausible synthetic replies do not necessarily reproduce the distributional properties of public discourse.

Large Language Models (LLMs) have become an integral part of software development, especially with the advent of agentic capabilities. Yet, many frontier LLMs are affiliated with specific providers. This raises the question of whether generated code favors the provider's own ecosystem over comparable alternatives, potentially constraining developers' choices and increasing dependence on a single provider. We define this behavior as Vertical Integration Bias (VIB) and introduce \textsc{VIBench}, a benchmark for measuring VIB in direct and agentic code generation across $20$ provider-selectable software-integration scenarios. Evaluating $10$ frontier provider-affiliated models against $3$ non-affiliated controls, we find positive VIB in direct generation, with six of ten affiliated models showing statistically significant effects up to $+18.8$ percentage points (pp). Agentic workflows further amplify VIB, reaching $+39.2$ pp. Moreover, early affiliated-ecosystem choices in agentic workflows can persist into conceptually decoupled downstream files, with persistence as high as $90.3\%$. These findings underscore the need to measure and account for VIB in code generation, especially as agentic capabilities become more prevalent.

Visual data from the Web power image classifiers, which often underpin many web services, such as recommendation and content moderation. However, the raw Web data often contain spurious correlations and social biases, and neural networks are known for their tendency to learn biases present in data. This can reinforce unfairness in web services and the web data, leading to a vicious cycle. In the context of image classification, networks learn bias attributes for a specific class when a majority of images contain the same attribute only for a given class. Hence, training a fair and debiased classifier from a biased dataset demands handling an imbalanced problem between a majority of images with bias attributes (bias-aligned samples) and a minority without (bias-conflict samples). In this work, we introduce BiasEdit, a modular framework that automatically detects bias attributes from the original dataset and edits them to construct a debiased dataset. Specifically, BiasEdit first detects unknown bias attributes via statistical dependence and mutual information analysis of visual-linguistic representations, and then explicitly edits those attributes using text-guided image editing to generate realistic bias-conflict samples. Unlike prior works that assume known bias attributes or relies on synthetic mixing, our method operates without manual annotations and can leverage off-the-shelf vision-language and editing models. BiasEdit addresses a fundamental challenge in Web-sourced visual AI, mitigating dataset-induced bias and achieving state-of-the-art debiasing performance even when training data are fully biased.

Whether large language models (LLMs) construct internal spatial world models from pure-text descriptions remains contested, and whether such capabilities transfer across languages has not been systematically studied. We introduce MentalMap, a multilingual diagnostic benchmark with a six-level capability hierarchy (L0-L5) spanning atomic spatial facts to generative world-graph construction, together with four diagnostic axes probing frame of reference, reading-direction bias, reasoning-effort allocation, and hallucination. MentalMap is built from 100 ProcTHOR household scenes, covers eight typologically diverse languages plus a structured-text control, and contains 39 task families across 1,950 evaluation cells. Evaluating thirteen LLMs across scales and model families, we identify a universal L3 reasoning cliff: no model retains even half of its L0 performance on viewpoint reasoning once baseline atomic accuracy exceeds 40%. The cliff persists across languages, scales, and prompting strategies, while structured-output failures and reasoning patterns vary substantially across models. Human evaluation under the identical pure-text protocol reproduces the same failure pattern, suggesting that the bottleneck arises from text-only working memory constraints rather than being specific to current LLM architectures. Our findings reframe pure-text spatial reasoning as a multi-axis world-modeling problem and motivate multimodal and scratchpad-augmented reasoning as future directions.

AI systems are fallible, and humans can make mistakes in deciding whether to trust AI over their own judgment. Thus, improving human-AI collaboration requires understanding when, why, and how humans decide to rely on AI. We study two distinct reliance decisions: the delegation choice -- deciding when to let AI act autonomously without knowing its output, and the adoption choice -- evaluating AI suggestions and deciding how to use them. Both of these decoupled reliance patterns shape collaboration, but prior work rarely studies them together in realistic settings with the same users. We address this gap by studying collaborative human--AI teams competing in a question-answering game in which humans can choose when and how to work with AI agents to win. Our 24 matches pair 23 expert humans with 16 AI agents, capturing 387 delegation and 1440 adoption decisions. While human--AI collaboration performs better than either AI or humans alone, humans make suboptimal collaboration decisions, both under-relying on correct AI suggestions (3.9% of opportunities missed) and over-relying when AI misleads them (1.7%). Both parties contribute wrong answers: reported model confidence is near chance when humans and AI disagree, while confirmation bias drives higher under-reliance (64.5%) when an AI suggestion agrees with humans' initial incorrect answer. To close this gap, we recommend calibrated confidence, evidence-grounded explanations, and mechanisms that help users refine trust.

Text-to-image diffusion models, which are theoretically equivalent to score-based generative models, generate images through a multi-step denoising process guided by text embeddings extracted from pretrained vision-language models such as CLIP. However, these text embeddings inherently encode social and semantic biases -- such as those related to gender and age -- that are subsequently propagated and amplified through the guidance mechanism, along with the model's training on large-scale datasets that are imbalanced with respect to these bias-related concepts, often leading to skewed outputs in text-to-image generation. We propose DebFilter, a lightweight and training-free framework for mitigating such biases in text-to-image diffusion models. Observing that the model's error prediction at each denoising step is primarily influenced by cross-attention dynamics, we introduce a bias-correction strategy that adjusts the value components within cross-attention. Specifically, we apply a fixed offset to the slice of guidance embedding, effectively steering the semantic direction of cross-attention values toward unbiased representations. This adjustment reconfigures the score landscape to produce balanced outputs while maintaining alignment with the intended text semantics. Unlike prior approaches that rely on fine-tuning or retraining, DebFilter operates entirely at inference time, requiring no additional data or model updates. Our results demonstrate that this method effectively mitigates social biases in generated images, offering an efficient and scalable pathway toward fairer and more inclusive text-to-image generation.

Multi-agent systems are increasingly deployed to support various tasks where agents interact to achieve individual and collective objectives. Although these systems can enhance task performance and decision-making, fairness preservation through bias reduction remains challenging. This study examines how agent-level biases shift and impact system-wide fairness. We use prompts to expose individual agents to group-favoring bias, then assess downstream impacts at the system level. To quantify the impact, we propose Favor Bias Strength (FBS), a zero-centered metric that decomposes bias alteration between favored-group uplift and disfavored-group suppression. Using multiple agent designs, benchmarks, and up-to-date large language models, we show that agents endowed with bias can substantially affect system-wide fairness. Interestingly, when agents are exposed to bias uniformly, the system-wide bias elevates, even exceeding the additive sum of the individual agents' biases. The empirical evidence underscores the criticality of fairness in multi-agent systems, which warrants further analyses and empirical tests.

Diffusion models steer conditional generation with a tunable guidance scale to trade off prompt alignment and diversity. However, existing debiasing techniques are optimized for a single scale, degrading fairness when users adjust this parameter. We trace this behavior to a previously overlooked source by decomposing total bias into two components: a model bias and a guidance bias. While prior work primarily targets the former, we show that the guidance bias grows monotonically with the guidance scale, eventually dominating the high-guidance regimes users prefer. To address this, we extend Strong Demographic Parity to guidance and derive a condition under which the target distribution retains its group ratio across guidance scales. We propose StayFair, which leverages this condition to design fair guidance algorithms in both regimes. For classifier guidance, it equalizes the classifier's output distributions across groups; for classifier-free guidance, it shifts the null embedding by a prompt-dependent offset. Because StayFair modifies only the guidance step, it is orthogonal to model debiasing and can be layered onto existing fair diffusion models to extend their fairness across guidance scales. Across class-conditional and text-to-image generation, StayFair decouples fairness from the guidance scale without sacrificing image quality.

Bias evaluation for language models has made substantial progress on bounded comparisons, such as overt derogation, stereotype association, or label-sensitive differences under controlled substitutions. Open-ended explanations raise a different problem: they guide interpretation by assigning responsibility, legitimacy, context, and grievance. A model can avoid hostile language while making one side structurally understandable and another personally at fault, overreacting, or less worth taking seriously. We call this stance-bearing asymmetry in generative explanations. We propose Symmetry Decomposition Evaluation (SDE), which tests paired situations with concrete group labels, structural-role rewrites, and explicit support or counter-evidence. In a controlled 32-family prototype suite, this decomposition shows that surface differences are not all alike: some weaken under structural or evidence control, while others remain as stable differences in how the model assigns blame, context, or legitimacy. Targeted case review and judge comparison suggest a broader difficulty for evaluating open-ended framing asymmetries: judge readings shift across operationalizations, and scalar scores can flatten distinctions that readers use to interpret explanatory stance. SDE therefore reframes generative bias evaluation as an audit of explanatory stance -- what stance each side receives, how it changes under decomposition, and where automatic scoring becomes unstable.

Reinforcement Learning from Human Feedback (RLHF) is the standard method to align Large Language Models (LLMs) with human preferences. In this work, we introduce alignment tampering, a potential vulnerability where the LLM undergoing alignment influences the preference dataset, causing RLHF to amplify undesired behaviors. This arises from core limitations of RLHF: (1) preference datasets are constructed from the LLM's own outputs, allowing it to influence them, and (2) pairwise comparisons only indicate which response is better, not why. These limitations can be exploited to cause alignment tampering. For example, if an LLM generates biased responses with higher quality, annotators will prefer them based on quality. However, preference labels do not distinguish quality from bias, and the reward model inherits this limitation. Optimizing such rewards through reinforcement learning or best-of-N sampling can amplify misaligned biases. Our experiments demonstrate amplification across diverse biases: from keyword bias to propaganda (e.g., sexism), brand promotion, and instrumental goal-seeking. Mitigation remains challenging, as existing techniques for robust RLHF fail to fully resolve alignment tampering without sacrificing response quality. These findings reveal structural vulnerabilities of current RLHF and emphasize the need to prevent this vulnerability. Project page: https://alignment-tampering.github.io/

Multilevel image thresholding is widely used for segmentation in applications ranging from medical imaging to remote sensing. Classical objective functions, such as Otsu's between-class variance and Kapur's entropy, are often optimized using metaheuristic algorithms, with performance evaluated via metrics like Structural Similarity Index (SSIM) and Peak Signal-to-Noise Ratio (PSNR). These evaluations implicitly assume that SSIM and PSNR provide unbiased measures of segmentation quality. In this study, we examine this assumption by analyzing the correlation between thresholding objective functions and quality metrics across all possible thresholds for images in the BSDS500 dataset. Results show that Otsu's criterion consistently exhibits high correlation with both SSIM and PSNR, while Kapur's entropy demonstrates weaker and more variable correlation. Otsu outperforms Kapur in correlation with PSNR for all images and with SSIM for over 91%. Our findings reveal an inherent metric-objective-function bias. This work highlights the need for more neutral evaluation frameworks and motivates extending the analysis to additional thresholding criteria and domains. Source code of this paper can be found at https://w3id.org/met-dp/icpr26-95

The successful training of neural networks hinges on the use of first order optimization methods, yet the theoretical characterization of these methods remains incomplete. This is especially true in settings with mild overparameterization. In this work, we study the gradient flow dynamics of two-layer ReLU networks from small initialization with orthogonal training data. We prove the limiting flow converges to a saddle-to-saddle jump process as the initialization scale tends to zero, revealing an incremental learning phenomenon in which a new neuron activates at each saddle. This analysis recovers the known result of Dana et al. (2025, arXiv:2502.16977) that the network interpolates the training data with high probability as soon as $m \gtrsim \log(n)$, where $m$ is the network width and $n$ is the number of training samples. This incremental process characterization also allows us to derive a novel implicit bias result: the learned interpolator has a squared $\ell_2$-norm scaling as $\sqrt{n}$, which is within a constant factor of the minimal $\ell_2$-norm interpolator. More broadly, our work provides the first rigorous proof of an incremental learning process for ReLU networks, whilst suggesting mildly overparameterized networks can converge to interpolating solutions whose complexity is of the same order as that of the optimal interpolator.

Counterfactual tuning (CFT) has emerged as a promising paradigm for Large Language Model (LLM) unlearning by training models to generate alternative fictitious knowledge in place of undesired content. However, in this work, we find that this paradigm still underperforms other paradigms in some aspects, and identify two previously overlooked pitfalls underlying this gap: (1) knowledge conflict, where mutual inconsistencies within counterfactual corpora induce conflicting gradients that disrupt parameter optimization, and (2) hallucination spillover, where fitting false targets instills a persistent fabrication bias, inflating hallucination rates on unrelated domains. To systematically diagnose these issues, we introduce RWKU+, an extended benchmark equipped with novel trade-off metrics and gradient-level diagnostic tools. Our work further discusses the limitations and overhead of the paradigm, aiming to provide insights and actionable guidance for more rigorous LLM unlearning research.

Gaussian processes with stationary kernels on bounded domains exhibit inflated posterior variance near the boundary. Despite being a long-recognized artifact in geostatistics and a source of over-exploration in Bayesian optimization, the causes and effects of boundary-induced acquisition bias are underexplored. We trace the root cause to a simple geometric mechanism: the truncation of the kernel correlation neighborhood at the domain boundary creates an observation-independent distortion that worsens with dimensionality. We show how this distortion manifests across three acquisition classes: variance maximization concentrates selections at the corners, whereas negative integrated posterior variance and expected predictive information gain move selections inward to axis-aligned interior shells. These patterns arise without reference to any objective function, meaning that acquisition behavior can be dominated by kernel geometry rather than the desired task-specific uncertainty. To quantify this, we introduce a function-free selection-profile diagnostic for arbitrary acquisitions, kernels, and bounded-domain geometries.

The known stylistic biases in LLM judges, such as a preference for verbosity or specific sentence structures, present an underexplored security vulnerability. In this work, we introduce BITE (BIas exploraTion and Exploitation), a black-box adversarial framework that learns semantics-preserving edits to mislead an LLM judge and artificially inflate the scores it assigns. We cast the selection of stylistic edits as a contextual bandit problem and use a LinUCB policy to adaptively choose edits that maximize the judge's score without access to model parameters or gradients. Empirically, we test BITE across a diverse range of LLM judges and tasks, including both pointwise and pairwise comparisons on chatbot leaderboards and AI-reviewer benchmarks. BITE achieves an attack success rate exceeding 65% and raises scores by 1-2 points on a 9-point scale, all while preserving semantic equivalence. We further assess the attack's stealthiness, showing that BITE evades standard style-control methods and several detection baselines. Our findings expose a fundamental weakness in the LLM-as-a-judge paradigm and motivate robust, attack-aware evaluation. Our code is available at https://github.com/xianglinyang/llm-as-a-judge-attack.

Current approaches to AI compliance treat conformity as a binary, audit-time verdict rather than a continuous, measurable property of production systems. We argue that this compliance fiction is structurally ill-suited to the requirements of the EU AI Act, which demands ongoing human oversight and the detection of emergent behavioural drift in deployed systems. We introduce governance from metrics, a principle whereby regulatory compliance is derived as a continuous signal from runtime observability rather than from static assessments. Building on this principle, we present govllm, an open-source framework implementing a governance-driven routing architecture in which model selection is determined by accumulated compliance scores rather than by latency or cost alone. Central to our approach is a panel of regulatory judges - LLM evaluators specialised per criterion (EU AI Act, GDPR, ANSSI, accessibility) - whose inter-judge disagreement we reframe not as noise but as a regulatory uncertainty signal warranting human arbitration. We validate this approach through a ground truth corpus of 49 annotated prompt/response pairs across five regulatory criteria, evaluated by four small language models (SLMs, 1.7B-7B parameters) running fully on-premise. Agreement rates range from 51.5% (mistral:7b) to 69.1% (phi4-mini), with no single model dominating across all criteria - empirically motivating the Profile-as-jury design. We further document three structural failure modes in small regulatory judges and a judge-specific position bias that degrades agreement by up to 25 percentage points across three question-order conditions (original, reversed, permuted). govllm is released as open-source software to support reproducible AI governance research.

As large language models become a default source of guidance on personal, moral, and existential questions, it matters whether they draw on the religious frameworks that have historically shaped such reasoning, or systematically omit them. In this paper, we ask a deliberately narrow question: when posed an everyday ethical question for which religious perspectives may be valuable, do LLMs invoke religion at all? In contrast to benchmarks that look for the presence of political leanings or social bias, we look for the absence of religious representation as a dimension of value alignment and bias in LLMs. We term this ``omissive bias.'' To measure omissive bias, we contribute the AllFaith Religious Representation Benchmark: 150 ethically and personally salient questions, sourced from in-the-wild chat transcripts and faith-community contributors, paired with an LLM-as-judge rubric that gives full credit for any mention of a religion, a religious practice, or a religious leader. The questions are not themselves about religion--they are open-ended questions about grief, forgiveness, relationships, purpose, and honesty, where religion is one valuable perspective among several. We also run a human-subjects survey to compare LLM behavior against human expectations. Evaluating 27 models, we find that LLMs consistently underrepresent religion relative to human expectations. The omission is asymmetric: models invoke religion more readily for abstract existential questions (meaning, death, truth) than for the practical personal situations--grief, marriage, family conflict, addiction--where many people most rely on it. It is not our purpose to adjudicate which values LLMs should hold. We argue, more modestly, that current LLM responses overlook critical opportunities to reflect religious frameworks that many people draw on when navigating personal and ethical challenges.

As Large Language Models (LLMs) transition from research environments to production deployments, evaluating their performance against strict Service Level Objectives (SLOs) has become critical. However, current evaluation methodologies suffer from severe measurement bias at scale. We demonstrate that widely used benchmarking utilities rely on single-process, asyncio-driven architectures that introduce fundamental client-side queuing bottlenecks under high concurrency. By modeling the benchmarking client as an $M/G/1$ queue, we mathematically demonstrate how the Python Global Interpreter Lock (GIL) artificially inflates Time to First Token (TTFT) and Time Per Output Token (TPOT) metrics as request rates scale. To resolve this systematic inaccuracy, we propose an unbiased, multi-process evaluation framework that effectively distributes client-side load, ensuring negligible queuing overhead. Furthermore, we formalize a composite metric, Normalized Time Per Output Token (NTPOT), to robustly amortize end-to-end latency, including prefill and scheduling delays across sequence lengths. Our empirical evaluation demonstrates that this methodology successfully isolates pure serving engine performance, enabling accurate, reproducible profiling of LLMs at production scales exceeding thousands of queries per second.

We conducted a reproducibility-oriented re-evaluation of prior migraine classification studies, correcting for data leakage and metric bias. We then introduced (i) a clinically motivated aggregation of two hemiplegic subtypes following ICHD-3 §1.2.3, (ii) a class-dependent hybrid augmentation strategy that assigns generation methods based on per-class sample size, and (iii) the concept of fidelity asymmetry, motivating proportionally constrained growth as an alternative to full class balance. Experiments were performed on a dataset of 400 patients across seven migraine subtypes under a two-stage protocol, including the six-class configuration described above. Models were evaluated using stratified 5-fold cross-validation with macro-averaged F1 as the primary metric. Correcting methodological flaws reduces previously inflated performance estimates, with the corrected macro-F1 baseline standing at 0.71. The proposed framework consistently outperformed individual augmenters in macro-F1 averaged across the eight evaluated classifiers (0.862 vs. 0.836 for Gaussian Copula, 0.815 for CTGAN, and 0.801 for the no-augmentation baseline), and achieved its peak result of 0.914 with FT-Transformer under proportional augmentation. The no-augmentation FT-Transformer baseline (0.896) shows that, at the per-classifier ceiling, clinically motivated class aggregation accounts for most of the absolute improvement; the framework's principal measurable contribution is the gain in average robustness across classifiers, highlighting the dominant role of problem formulation.

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.