inference-time scaling

Why LLMs Generalize on Maps but Fail on Scale

A synthetic shortest-path setup shows LLMs transfer across maps, but break when problems get longer because recursive reasoning gets unstable.