The Flow has not summarised this story yet — read it at arXiv Computation and Language.
BiasGym is a cost‑effective, generalizable framework that injects specific biases into large language models via token‑based fine‑tuning while keeping the model frozen. It then uses two debiasing methods—Scope and Steer—to identify and suppress or redirect the components responsible for biased behavior. The framework enables consistent bias elicitation, precise localization of bias associations, and targeted debiasing without harming downstream performance, and it has been shown to reduce real‑world stereotypes such as labeling Italians as reckless drivers.
arXiv:2311.13892v4 Announce Type: replace-cross Abstract: The social biases and unwelcome stereotypes revealed by pretrained language models are becoming obstacles to their application. Compared to n...
arXiv:2608. 14161v1 Announce Type: new Abstract: LLMs exhibit social biases that can produce inaccurate and discriminatory inferences, posing risks in high-stakes applications.
The paper presents GRASP, a method for unsupervised detection and removal of spurious correlations that arise during fine‑tuning of pretrained language models. By identifying latent factors from LoRA fine‑tune weights and projecting gradients to prevent new reliance on them, GRASP preserves genuine task signals while mitigating bias. Experiments on three tasks—misaligned code generation, medical advice, and political bias—show that GRASP reduces misalignment and drift more effectively than existing baselines while maintaining task performance.
The paper examines how large language models (LLMs) can be biased by irrelevant social contexts when evaluating teachers, using a large U.S. classroom transcript dataset. It shows that spurious contexts can shift model ratings by up to 1.48 points on a 7‑point scale and that standard mitigation methods like SFT and DPO are insufficient. The authors introduce Debiasing‑DPO, a method that combines contrastive reasoning‑augmented DPO with SFT, which reduces bias by 84% and improves predictive accuracy by 52% on Llama and Qwen Instruct models.
The paper introduces JudgeBiasBench, a benchmark that systematically quantifies judgment biases in large language model (LLM)-based judges across four dimensions and 12 bias types. It evaluates both generative and discriminative judges, revealing significant bias patterns that undermine reliability. The authors propose bias-aware training—reinforcement learning for generative judges and contrastive learning for discriminative judges—to reduce these biases while maintaining evaluation performance.