BiasGym: A Simple and Generalizable Framework for Analyzing and Removing Biases through Elicitation
Understanding biases and stereotypes encoded in the weights of Large Language Models (LLMs) is crucial for developing effective mitigation strategies. However, biased behaviour is often subtle and non-trivial to isolate, even when deliberately elicited, making systematic analysis and debiasing particularly challenging. To address this, we introduce \texttt{BiasGym}, a simple, cost-effective, and generalizable framework for reliably and safely injecting, analyzing, and mitigating conceptual associations of biases within LLMs. \texttt{BiasGym} consists of two components: \texttt{BiasInject}, which safely injects specific biases into the model via token-based fine-tuning while keeping the model frozen, and \texttt{BiasScope}, which leverages these injected signals to identify and reliably steer the components responsible for biased behavior. Our method enables consistent bias elicitation for mechanistic analysis, supports targeted debiasing without degrading performance on downstream tasks, and generalizes to biases unseen during fine-tuning. We demonstrate the effectiveness of BiasGym in reducing real-world stereotypes (e.g., people from Italy being `reckless drivers'), showing its utility for both safety interventions and interpretability research.
Code (0)
등록된 구현이 없습니다.
Similar Papers 제목 키워드 기반
Technical Report: Quantifying and Analyzing the Generalization Power of a DNN
This paper proposes a new perspective for analyzing the generalization power of deep neural networks (DNNs), i.e., directly disentangling and analyzing the dynamics of generalizable and non-generalizable interaction enco…
MILO: Model-Agnostic Subset Selection Framework for Efficient Model Training and Tuning
Training deep networks and tuning hyperparameters on large datasets is computationally intensive. One of the primary research directions for efficient training is to reduce training costs by selecting well-generalizable …
modelDebiased Learning for Remote Sensing Data
Deep learning has had remarkable success at analyzing handheld imagery such as consumer photos due to the availability of large-scale human annotations (e.g., ImageNet). However, remote sensing data lacks such extensive …
SemDeDup: Data-efficient learning at web-scale through semantic deduplication
Progress in machine learning has been driven in large part by massive increases in data. However, large web-scale datasets such as LAION are largely uncurated beyond searches for exact duplicates, potentially leaving muc…
Breaking Down Bias: On The Limits of Generalizable Pruning Strategies
We employ model pruning to examine how LLMs conceptualize racial biases, and whether a generalizable mitigation strategy for such biases appears feasible. Our analysis yields several novel insights. We find that pruning …
Decision Making