BiasTrace: Linking Reasoning Behaviours to Biased Outputs in LLMs

2026-08-14

AI Summary

This study addresses the prevalent oversight of reasoning mechanisms and the difficulty in tracing bias origins within large language models. We propose BiasTrace, a pioneering reasoning-level bias annotation framework that correlates reasoning behaviors with biased outputs. By integrating LLM-as-a-Judge evaluation with inference-time intervention strategies, this approach elucidates the primary causes of implicit reasoning-induced bias. Extensive experiments on a newly constructed large-scale dataset demonstrate that reasoning-level annotations not only significantly enhance bias detection accuracy but also effectively improve mitigation performance during the inference phase. Consequently, this work establishes a novel paradigm for both understanding and governing model bias by shifting focus from output-level symptoms to underlying reasoning processes.

Abstract

LLMs exhibit social biases that can produce inaccurate and discriminatory inferences, posing risks in high-stakes applications. While prior work has made progress in measuring and mitigating bias, it largely focuses on final outputs of models, with limited understanding of the mechanisms that produce biased outcomes. Recent advances in LLM reasoning offers a new lens for investigating bias, yet the link between reasoning and bias remains poorly understood. Existing approaches focus primarily on final answer correctness or explicitly biased language, overlooking different behaviours in reasoning that can drive biased outcomes. We introduce BiasTrace, an annotation scheme for labelling reasoning behaviours in model-generated traces and linking them to biased outcomes. BiasTrace captures bias-specific behaviours (e.g., unsupported demographic assumptions) as well as general reasoning patterns that may implicitly contribute to bias (e.g. overthinking). We apply BiasTrace to reasoning traces in bias-sensitive contexts, scaled using validated LLM-as-a-judge methods, producing a large annotated dataset. Our analysis shows that biased outputs often stem from subtle reasoning behaviours rather than explicitly biased language, and that reasoning-level annotations improve bias detection. We further show that BiasTrace behaviours can be exploited for inference-time mitigation. These findings underscore the importance of examining a broader range of reasoning patterns to better understand bias in LLMs.

Problem

• LLM bias
• reasoning behaviors
• bias mechanism
• reasoning traces

Innovation

• BiasTrace
• Reasoning Behaviours
• LLM-as-a-judge
• Inference-time Mitigation
• Bias Detection