I've noticed there's some stochasticity around this too. I think this self-correcting behavior in the middle of a response is among my least favorite. It should resolve this kind of issue with internal reasoning.
AI Self-Correction Behavior During Response Generation Issues
By
–