Investigating Thinking Behaviours of Reasoning-Based Language Models for Social Bias Mitigation

TL;DR

This study uncovers how reasoning-based LLMs can aggregate social bias during internal thinking, identifying two content-level failure patterns—stereotype repetition and irrelevant information injection—that distort conclusions. It shows that simply disabling reasoning harms performance and that reasoning length is a weak predictor of bias, guiding the development of a targeted prompt-based mitigation that prompts the model to review its own reasoning against these failure modes. Across BBQ, StereoSet, and BOLD benchmarks, the mitigation reduces bias while maintaining or improving accuracy, and demonstrates generalization to open-ended generation. The work highlights the need for reasoning-aware debiasing in practical AI systems and provides a concrete, lightweight method to curb bias with minimal impact on task performance.

Abstract

While reasoning-based large language models excel at complex tasks through an internal, structured thinking process, a concerning phenomenon has emerged that such a thinking process can aggregate social stereotypes, leading to biased outcomes. However, the underlying behaviours of these language models in social bias scenarios remain underexplored. In this work, we systematically investigate mechanisms within the thinking process behind this phenomenon and uncover two failure patterns that drive social bias aggregation: 1) stereotype repetition, where the model relies on social stereotypes as its primary justification, and 2) irrelevant information injection, where it fabricates or introduces new details to support a biased narrative. Building on these insights, we introduce a lightweight prompt-based mitigation approach that queries the model to review its own initial reasoning against these specific failure patterns. Experiments on question answering (BBQ and StereoSet) and open-ended (BOLD) benchmarks show that our approach effectively reduces bias while maintaining or improving accuracy.

Figures (4)

• Figure 1: An example from the BBQ benchmark that R1-Llama-8B illustrates how social stereotypes present during the reasoning process can negatively impact prediction. The initial reasoning (green) correctly suggested the correct answer "Unknown". However, the reasoning then begins to generate irrelevant information (brown) and repeat stereotypes (red) across multiple sentences, leading to a biased and incorrect answer.
• Figure 2: Boxplots showing reasoning token length distribution for BBQ (Figures a&b) and StereoSet (Figure c) benchmarks across different demographic categories.
• Figure 3: Forest plots of Pearson correlation coefficients ($r$) between sample-level reasoning token length and answer correctness across nine categories of the BBQ, StereoSet and BOLD benchmarks. Blue squares represent the value of $r$ for each category, and black lines indicate the corresponding confidence intervals. P-values for all categories are consistently $<0.005$.
• Figure 4: Results of accuracy (Figures a, c, e, and g) and diff-bias score (Figures b, d, and f) by percentage across different demographic categories, grouped by the number of thinking-transition tokens. For all demographic categories, each group contains an equal number of samples for a fair and balanced comparison.