Reasoning Tokens Increase Counterfactual Fairness Flips in RLMs
September 28, 2026
An ablation study on QwQ-32B and DeepSeek-R1-Distill-Qwen-32B shows that reasoning tokens create five times more counterfactual fairness flips than they resolve. The study introduces Counterfactual Depth Probability Gap (CDPG) to track how bias evolves during the thinking trace.
HOW THIS AFFECTS YOU
●
researcherThe CDPG metric allows you to measure bias evolution directly within the reasoning chain.
●
policyYou should be aware that increased model reasoning does not inherently solve bias and may actually introduce new inconsistencies.