Spending the same thinking budget on every problem backfires — too much reasoning on easy questions actually tanks accuracy.

This explores what accuracy you actually gain by matching thinking budget to each problem's difficulty, rather than spending the same number of tokens on every problem. The short version from the corpus: the gains come less from thinking *more* and more from thinking the *right amount* — and a fixed budget is almost always the wrong amount for most problems. The core finding is that the token budget maximizing accuracy differs by difficulty, so any single fixed number simultaneously over-thinks easy problems and under-thinks hard ones Does thinking time need to scale with problem difficulty?. That's not a small effect: pushing thinking tokens from ~1,100 up to ~16K dragged benchmark accuracy *down* from 87.3% to 70.3%, because extra reasoning on easy problems curdles into second-guessing Does more thinking time always improve reasoning accuracy?.

So the honest framing is that 'adaptive vs fixed' isn't adaptive winning by a few points — it's fixed budgets leaving accuracy on the table in two directions at once. When you hold *total* compute constant and just redistribute it — less for easy prompts, more for hard ones — you outperform even larger models running under uniform budgets Can we allocate inference compute based on prompt difficulty?. The lever is allocation, not volume. This shows up as a shape, too: accuracy against reasoning length traces an inverted U, where the peak shifts right for harder tasks and left for more capable models Why does chain of thought accuracy eventually decline with length?. A fixed budget picks one point on that curve for everything; adaptation tracks the peak as it moves.

Where it gets more interesting is that adaptation can be baked in at different stages. During *training*, budgets that start generous and gradually tighten beat fixed-budget training — the model first explores strategies with room to breathe, then learns to compress them, ending up both more accurate and more token-efficient Does gradually tightening token budgets beat fixed budget training?. At *inference*, a model can learn to route itself: Thinkless trains one model to choose between extended reasoning and a direct answer, self-calibrating when a problem even warrants thinking — no difficulty labels required Can models learn when to think versus respond quickly?. And you can even steer verbosity after the fact, since concise and verbose reasoning occupy different regions of activation space — one extracted vector cut chain-of-thought length 67% while holding accuracy Can we steer reasoning toward brevity without retraining?.

The thing you might not expect: whether extra thinking helps at all depends on the *training regime*, not the budget knob. Vanilla models use thinking mode counterproductively — more tokens means more self-doubt — until RL training flips that same mechanism into productive gap analysis Does extended thinking help or hurt model reasoning?. That's why non-reasoning models can't simply be handed a bigger budget to catch up; without a trained reasoning protocol, additional tokens aren't productive Can non-reasoning models catch up with more compute?. Adaptive budgeting is what lets a model that *can* reason spend its allowance well; it doesn't manufacture the ability to reason in the first place.

So the accuracy gain from adaptive budgets is real and sometimes large, but it's best understood as recovering the accuracy that fixed budgets waste — clawing back the points lost to overthinking easy items and underthinking hard ones — rather than as a new capability. The frontier work isn't 'how many tokens' but 'who decides, and when': curriculum during training, self-routing at inference, or activation steering after generation.