When a mind — artificial or otherwise — circles the same ground repeatedly, it is not reasoning but ruminating. Researchers have now given large language models a way to recognize their own loops: a lightweight framework called ISST that watches the model's internal uncertainty in real time and nudges it toward new paths when it grows too comfortable with repetition. Tested across major mathematical benchmarks, the approach cuts repetitive generation by half without touching the model itself — a reminder that sometimes the wisest intervention is not rebuilding the thinker, but quietly adjustin