Large Reasoning Models' Overthinking Reduces Accuracy by Up to 21%

A new study from Simone Caldarella and colleagues at the University of Trento and Bosch investigates a critical flaw in Large Reasoning Models (LRMs): harmful overthinking. While LRMs improve performance by generating explicit intermediate reasoning traces, the assumption that more reasoning is always better is challenged. The researchers introduce a prefix-level trajectory evaluation protocol that defines the minimum reasoning budget needed for a model to first produce the correct answer. They distinguish between verbose overthinking (redundant but harmless) and harmful overthinking (where additional reasoning destabilizes an already correct trajectory).

On multimodal benchmarks, the team found that many reasoning-intensive instances require surprisingly little reasoning. Stopping at the first correct prefix boosts accuracy by up to 21% compared to standard full reasoning. However, common efficiency strategies like early stopping—which cut verbose overthinking by 50%—fail to mitigate harmful overthinking. Failure analysis reveals that correctness deviations are driven by logical drift and visual reinterpretation. The results generalize to language-only reasoning benchmarks, highlighting harmful overthinking as a broader reliability risk for AI systems.