Adaptive Triggering Uses Change-Point Detection to Correct LLM Reasoning Bias

This method treats bias mitigation in chain-of-thought reasoning as an online change-point detection problem using a CUSUM statistic. By injecting corrections only when accumulated bias signals cross a calibrated threshold, the approach prevents biased reasoning from propagating without disrupting correct reasoning trajectories.