AI reasoning models now detect their own mistakes mid-thought and fix them without retraining

Researchers identified specific neurons in large reasoning models that fire when the model makes calculation errors, gets stuck in loops, or overthinks a problem. They built a lightweight correction system that detects these neuron patterns during inference and triggers the model to restart that step, improving accuracy by up to 27% while using 20-60% fewer tokens.

For years, when reasoning models failed, the failure was opaque — you couldn't see where thinking broke down or why. This work makes the failure visible at the neuron level and turns that visibility into a repair mechanism that works without retraining. The practical effect is simple: a reasoning model can now self-correct on the fly, which means you get better answers without running the model twice or retraining it on failure cases. The cost savings matter too — using 40-60% fewer tokens on corrected problems is the difference between an expensive system and a practical one.