Large reasoning models (LRMs) have shown strong results on challenging math and science problems by generating long chain-of-thought reasoning. However, this often leads to overthinking, where the model adds redundant or unnecessary steps that increase compute cost without improving accuracy.
To address this, researchers propose CALM (Consistency-Aware Logit Masking), a method that intervenes during decoding. Instead of modifying the model's training or prompting, CALM adjusts the logits at each generation step by masking options that are inconsistent with the reasoning path, encouraging the model to converge more efficiently.
The work is described in a new arXiv preprint, though details beyond the abstract are limited. The authors position CALM as a decoding-space solution, which could make it easier to apply to existing models without retraining. Further evaluation would be needed to confirm how much overthinking is reduced and whether performance on complex tasks remains intact.