LIDAO: Towards Limited Interventions for Debiasing (Large) Language Models

TL;DR

This work tackles global bias in large language models by linking bias to mutual information between a global property $g$ and a demographic attribute $a$. It shows that constant, per-token constraints are overly restrictive and introduces LIDAO, a limited-intervention framework that allows tokens to be relevant to either $g$ or $a$ as long as per-token independence holds, yielding better fairness-fluency trade-offs. A formal theorem establishes that sufficient per-token independence implies $I(g; a)=0$, and two practical variants (min-based and prod-based) enable adaptive, decoding-time debiasing using lightweight MI proxies. The approach is extended to adversarial prompts via eLIDAO, which uses a reference model to guide generation in a way that reduces prompt-driven bias while maintaining fluency. Experimental results across three models and three tasks show that LIDAO/eLIDAO outperform baselines in reducing bias and preserving text quality, with eLIDAO offering robust performance under adversarial prompts. Overall, the paper provides a principled, scalable path to debiasing large language models with limited interventions and demonstrates practical impact for safer and more fluent NLG systems.

Abstract

Large language models (LLMs) have achieved impressive performance on various natural language generation tasks. Nonetheless, they suffer from generating negative and harmful contents that are biased against certain demographic groups (e.g., female), raising severe fairness concerns. As remedies, prior works intervened the generation by removing attitude or demographic information, inevitably degrading the generation quality and resulting in notable \textit{fairness-fluency} trade-offs. However, it is still under-explored to what extent the fluency \textit{has to} be affected in order to achieve a desired level of fairness. In this work, we conduct the first formal study from an information-theoretic perspective. We show that previous approaches are excessive for debiasing and propose LIDAO, a general framework to debias a (L)LM at a better fluency provably. We further robustify LIDAO in adversarial scenarios, where a carefully-crafted prompt may stimulate LLMs exhibiting instruction-following abilities to generate texts with fairness issue appears only when the prompt is also taken into account. Experiments on three LMs ranging from 0.7B to 7B parameters demonstrate the superiority of our method.

Theorems & Definitions (10)

• Theorem 2.1
• Remark 2.2
• Proposition 3.1
• Remark 3.2
• Theorem
• Lemma 1.1
• proof : Proof of Lem \ref{['lem:lidao']}
• proof : Proof of Thm \ref{['thm:lidao']}
• Proposition
• proof