Why does more reasoning sometimes make models worse?

The dominant test-time-compute assumption is monotonic: more thinking, better answers. "When More Thinking Hurts" falsifies it with a specific mechanism. Tracking flip events — moments where the model, mid-trace, moves off an answer it had already reached — shows overthinking is not merely diminishing returns but active regression: extended reasoning is associated with abandoning previously-correct answers. Accuracy is therefore non-monotonic in trace length, not just concave.

This sharpens the vault's overthinking thread. Does more thinking time always improve reasoning accuracy? establishes that a threshold exists; the flip-event lens names what happens at it — the model second-guesses a correct commitment and talks itself out of it. It is distinct from Do reasoning models switch between ideas too frequently?: underthinking abandons a path too early before it pays off, whereas a flip event abandons a path that had already produced the right answer. Both are switching failures, but on opposite sides of the payoff.

The practical consequence is a measurement demand: report efficiency frontiers, not just accuracy-vs-budget curves. Because marginal utility falls sharply and can go negative, the paper's cost-aware framing shows stopping at a moderate budget can cut computation substantially while holding accuracy roughly constant — the extra tokens were buying flip risk, not correctness. This is a direct argument against uniform "think as long as possible" decoding.