Models can reach the right answer and then — by thinking longer — talk themselves into the wrong one.

This question asks what breaks when a model reasons past its optimal length — and the corpus points to a clear answer: the extra tokens don't stall out neutrally, they actively corrupt answers the model had already reached. The headline number is stark. As thinking tokens scale from ~1,100 to ~16,000, benchmark accuracy falls from 87.3% to 70.3% — a non-monotonic curve where accuracy peaks and then declines sharply rather than plateauing When does thinking too much actually hurt reasoning? Does more thinking time always improve reasoning accuracy?. So the degradation isn't diminishing returns; it's a reversal.

The mechanism is second-guessing. Tracking individual reasoning traces shows 'flip events' — the model reaches a correct answer, keeps going, and then talks itself out of it, overwriting the right answer with a wrong one Why does more reasoning sometimes make models worse?. Extended thinking inflates output variance and introduces self-revision errors When does thinking too much actually hurt reasoning?. A related failure is churn rather than reversal: models switch reasoning paths too early and too often, abandoning promising approaches mid-exploration and burning tokens on incomplete threads — penalizing those thought-transition tokens at decode time alone recovers accuracy with no retraining Do reasoning models switch between ideas too frequently?. Overthinking (flipping away from good answers) and underthinking (never committing to one) are two faces of the same missing skill: knowing when to stop.

That 'when to stop' skill is exactly what reasoning training tends not to install. The optimal chain-of-thought length follows an inverted-U — it rises with task difficulty but *shrinks* as the model gets more capable, and RL training naturally drifts toward shorter chains as competence grows, meaning brevity is something models earn from reward signals rather than something forced on them Why does chain of thought accuracy eventually decline with length?. When training instead optimizes purely for producing reasoning steps, models lose the ability to disengage: they generate long, redundant traces for ill-posed questions with missing premises that non-reasoning models correctly flag as unanswerable Why do reasoning models overthink ill-posed questions?, and more broadly they narrow their critical-thinking range even as they look more capable What critical thinking skills do reasoning models actually lose?.

Here's the part you might not expect: the reasoning tokens themselves may not be doing the work you'd assume, which reframes *why* extra ones can hurt. Models trained on deliberately corrupted or irrelevant traces stay just as accurate — traces act as computational scaffolding, not as meaningful step-by-step logic Do reasoning traces need to be semantically correct?. And intermediate tokens carry no special execution semantics; invalid traces routinely yield correct answers, so the trace correlates with the answer through learned formatting, not verified inference Do reasoning traces actually cause correct answers? Why does chain-of-thought reasoning fail in predictable ways?. If reasoning is largely pattern-matched scaffolding rather than genuine deliberation, then generating *more* of it past the point of usefulness is just more opportunity for the pattern to drift off the correct answer — degradation isn't a bug in the reasoning, it's what happens when there's no real reasoning holding the answer in place.

Two practical upshots. First, the threshold is real but invisible until crossed — it varies by model, task, and difficulty with no reliable predictor, though difficulty estimators and runtime confidence signals can start to detect it dynamically How can we predict the optimal thinking token threshold?. Second, the degradation is not inevitable: RL training can flip 'thinking mode' from counterproductive self-doubt into productive gap analysis, which means training mediates reasoning *quality*, not just quantity — the same extended-thinking mechanism that hurts a vanilla model can help a well-trained one Does extended thinking help or hurt model reasoning?.