When a language model keeps its reasoning trace at answer time, that trace does double duty. It can help the model arrive at a correct answer, but it can also stubbornly compete with fresh evidence when the two disagree. A new arXiv preprint sets out to separate these two effects rather than treating them as one.
The authors cross verified evidence versions with the traces, allowing them to isolate how much of the final answer comes from the trace's internal logic versus the external evidence. This design makes it possible to measure both the accuracy boost from retaining a trace and the cost when that trace points in a different direction from the current facts.
The work underscores a practical tension for model deployment: keeping reasoning traces can improve performance, but it may also make models less responsive to corrections or new information. Understanding when each role dominates could guide how traces are stored, truncated, or discarded in future systems.