fairness

The paper introduces INCLUDE, a multilingual evaluation benchmark to quantify Indian-centric socio-cultural biases in LLMs, revealing that non-English Indian languages exhibit higher bias than English, indicating cross-lingual safety alignment gaps.

This article demonstrates how high validation accuracy can conceal proxy bias in AI models, using SHAP for explainability and a runtime governance wrapper to enforce fairness policies and prevent biased decisions in production.

This position paper argues that fairness failures in generative models are primarily due to evaluation problems and proposes Fairness Cards as a standardized reporting artifact to improve reproducibility and accountability.

This paper proposes FairTFM, a training strategy to incorporate fairness constraints into tabular foundation models, enabling fair predictions via in-context learning without task-specific retraining.

This paper investigates whether judges in misdemeanor bail hearings behave like algorithms by training interpretable ML models on court data, finding that judges generally follow small formula-like rules but sometimes show substantial inconsistency and unequal treatment across similar defendants.

This paper identifies procedural fairness failures in RLHF caused by averaging heterogeneous preferences, where majority groups dominate reward learning and minority preferences are under-represented. It proposes Preference-Aware RLHF (PA-RLHF), which improves alignment accuracy and reduces the fairness gap in controlled experiments.

This paper audits confidence-based label filtering in automated credit default prediction, revealing income-conditional recall disparities and showing that blinding models to sensitive attributes does not eliminate fairness gaps when proxies and institutional biases persist.

UNMASK is an automated pipeline for discovering and causally verifying spurious shortcuts in text classifiers, enabling mitigation without human annotations and improving robustness on benchmarks.

This paper introduces a novel audit protocol using Private Information Retrieval to make audits manipulation-proof, forcing deceptive model providers to falsify more responses and increasing detection likelihood, with theoretical guarantees and experimental validation.

This paper studies the sample complexity of multicalibration for a sequence of properties that are sequentially identifiable, establishing matching upper and lower bounds up to logarithmic factors.

This paper introduces a constrained mixed-strategy GroupDRO framework for equitable system-prompt selection, assigning weights to existing prompts to minimize worst-case information-quality loss across demographic groups and metrics. Experiments across five LLMs on bilingual medical and finance benchmarks show consistent reductions in worst-case quality drops while preserving average performance.

This paper introduces the 'fairness collapse' phenomenon, showing that training language models on synthetic data silently amplifies social biases before standard model collapse metrics degrade, highlighting a critical risk for AI fairness.

This paper proposes HEIMAT, a heuristic-style automatic debiasing framework for language models that uses heuristic prompts to reveal biases and fine-tunes the model to reduce bias while preserving NLU performance.

This paper proposes an inference-time policy shaping framework to steer pretrained reinforcement learning policies toward welfare-based fairness objectives without retraining, inspired by inference-time alignment in LLMs.

A study evaluating zero-shot GPT-4o-mini for predicting child stunting from Bangladesh Demographic and Health Survey data, comparing against a random forest baseline and assessing fairness across demographic groups and temporal robustness. Results show comparable balanced accuracy but notable fairness disparities across residence and wealth categories.

Introduces FairFund-Bench, a benchmark for evaluating distributive bias in LLM resource allocation, showing that audit format changes the direction and magnitude of bias, and that causal framing effects dominate demographic effects.

Introduces FairDiffuseVQVAE, a two-stage tabular diffusion model that achieves fairness at sampling time by conditioning on protected attributes, outperforming prior fair tabular generators on demographic parity and equalized odds.

This paper presents Fairness Pruning, a structural intervention method that locates demographic bias in GLU-MLP layers of large language models by identifying differentially activated neurons. Zeroing a very small number of neurons disrupts bias processing while preserving reasoning and general knowledge, suggesting bias and capabilities rely on dissociable circuits.

This paper introduces S2D, a systematic framework to evaluate regional bias in LLMs from abstract stereotypes to concrete social decisions, covering all 34 provincial-level regions of China. Results show that regional bias is prevalent, systematic, and consequential, with patterns linked to regional economic indicators.

This paper introduces Symphony-Bias, a multimodal dataset for evaluating gender bias in LLMs' associations with musical instruments across text, vision, and audio. It finds that 92% of instrument-level outcomes align with prior social-science findings, with gender biases strongest in text and weakest in audio.