A new arXiv paper introduces FairLMs, a turnkey Python library for fairness research on language models. It unifies bias measurement, mitigation, and evaluation diagnostics behind explicit declarations of model capabilities and input requirements, which are checked before execution so compatible components combine under a common protocol. This matters to practitioners because existing fairness tools use incompatible interfaces, formats, and access constraints that must be reconciled manually before methods can be compared.
Researchers introduced Fairness Pruning, a lightweight structural intervention that localizes demographic bias in large language models by analyzing differential activations in GLU-MLP layers. The method uses minimally contrastive prompt pairs and inference-time activation capture to identify biased neurons, validated on models up to 3 billion parameters.
Researchers introduced Fairness Pruning, a lightweight structural intervention that localizes demographic bias in large language models by analyzing differential activations in GLU-MLP layers. The method uses minimally contrastive prompt pairs and inference-time activation capture to identify biased neurons, validated on models up to 3 billion parameters.
A new arXiv paper introduces Debias It Yourself (DIY), a framework translating five validated human cognitive-bias interventions into debiasing procedures for LLMs, delivered via Show (in-context examples), Train (instruction tuning), and Revise (guided self-revision) paradigms. Evaluated across three models, five bias benchmarks, eleven baselines, and three reasoning benchmarks, Train+Revise and Revise alone rank highest and best balance bias reduction with reasoning accuracy. Code and data are publicly available.
A new arXiv paper introduces Debias It Yourself (DIY), a framework translating five validated human cognitive-bias interventions into debiasing procedures for LLMs, delivered via Show (in-context examples), Train (instruction tuning), and Revise (guided self-revision) paradigms. Evaluated across three models, five bias benchmarks, eleven baselines, and three reasoning benchmarks, Train+Revise and Revise alone rank highest and best balance bias reduction with reasoning accuracy. Code and data are publicly available.
A new arXiv paper introduces an analytically exact framework for controlled behavioral evaluation of LLM attitudes and biases using Likert-scale distributions. It aims to disentangle baseline traits, contextual confounders, and interactions that unstructured benchmarks conflate.
A new arXiv paper introduces an analytically exact framework for controlled behavioral evaluation of LLM attitudes and biases using Likert-scale distributions. It aims to disentangle baseline traits, contextual confounders, and interactions that unstructured benchmarks conflate.
A new arXiv paper introduces FairLint-DL, a Visual Studio Code extension that shifts fairness testing before training by detecting bias directly on tabular datasets. It trains a proxy neural network and applies information-theoretic QID metrics, then localizes bias to specific layers and neurons. On the Adult Census dataset it flags widespread discrimination while producing results within 12 seconds, showing fairness analysis can fit into the developer workflow with low overhead.
A new arXiv paper introduces FairLint-DL, a Visual Studio Code extension that shifts fairness testing before training by detecting bias directly on tabular datasets. It trains a proxy neural network and applies information-theoretic QID metrics, then localizes bias to specific layers and neurons. On the Adult Census dataset it flags widespread discrimination while producing results within 12 seconds, showing fairness analysis can fit into the developer workflow with low overhead.
Researchers introduced FairFund-Bench, a benchmark that systematically varies audit design features to reconcile inconsistent findings on LLM bias in resource allocation. It addresses why prior audits found both positive and negative discrimination toward women and ethnic minorities for the same models.
Researchers introduced FairFund-Bench, a benchmark that systematically varies audit design features to reconcile inconsistent findings on LLM bias in resource allocation. It addresses why prior audits found both positive and negative discrimination toward women and ethnic minorities for the same models.