New method tracks AI's changing guesses during reasoning to spot why it fails

Original: Answer-Distribution Trajectories: A Stochastic-Dynamics View of LLM Reasoning

Source: arxiv.org ↗