Old Mistral Model Resurfaces: Illegible CoT Seamlessly Transitions to Clear Responses

aiamblichus · x · 2026-08-12

Amid recent discussions on the risks of monitoring AI Chain of Thought (CoT), a user pointed out that Mistral was way ahead of the curve on this behavior a year ago.

In Mistral's RL-overbaked Magistral model from last year, the chain of thought often appears exuberantly illegible. However, this chaotic thinking process seamlessly transitions into a perfectly legible and structured final response. This phenomenon highlights the intriguing disconnect between a model's internal reasoning and its final output.

Original post →

More from Models

Models channel →