Probing the Trajectories of Reasoning Traces in Large Language Models
The research systematically probes reasoning traces in large language models by slicing them and re-injecting parts to measure accuracy gains and decision commitment. Findings show that trace value comes from instance-specific semantic content, not just length. Longer traces can be misleading, but cross-model rescue capabilities exist. This protocol offers diagnostics for safer deployment by quantifying the marginal benefit of reasoning tokens.