A reasoning model spends 901 tokens to answer 2 plus 3

Thirteen solutions to a two-second question

Ask a recent reasoning model a question with an instant answer and watch it think anyway. Chen and colleagues put the plainest possible version to QwQ-32B-Preview: what is the answer of 2 plus 3? The model produced 901 tokens of reasoning, spread across thirteen separate solution attempts, before it gave the five that a person reaches in two seconds. A conventional model answered in about thirty-nine.¹

One could reasonably assume that all of that visible deliberation is buying a surer answer. An earlier lesson, a transformer can only compute more by writing more, showed why writing intermediate steps is the only way a fixed network spends more computation on a problem. That spent computation is the model's test-time compute, the work it does to answer one prompt. For these models it rises directly with the number of tokens generated.

Which is what makes the thirteen attempts strange. The answer to 2 plus 3 was never in question, and the model spent a hard problem's worth of computation confirming it. That is one half of the behavior. The other half sounds like its opposite: on certain tasks, letting a model reason longer makes its answer worse. Both start from the same place, a model trained to reason at length.

The chain grew longer on its own during training

A reasoning model is not told how long to think. DeepSeek trained its R1-Zero model by pure reinforcement learning. The model was rewarded only for reaching the correct final answer on problems with checkable solutions, and it saw no human-written reasoning to imitate. Reinforcement learning here means exactly that and nothing more. Get the answer right, get a positive signal.

Two things happened over that training. Accuracy on hard math climbed steeply: on AIME 2024, a demanding high-school competition, the model's first-try success rate rose from 15.6% to 71.0% across the run.² And the chains grew. Without anyone setting a target length, the average answer stretched from hundreds of tokens to thousands as training went on.

The reward was a correct final answer. What reached correct answers on hard problems was longer deliberation, so longer deliberation is what the training selected. The result is a policy that, by default, thinks at length before it commits. Nobody wrote a rule that said deliberate for 900 tokens. The rule the model learned was reach the answer, and on the problems that shaped it, reaching the answer took a long chain.

No gauge tells it the question was easy

That policy is now fixed in the model, and it carries no reliable sense of how hard the current question is. There is no step where the model sizes up 2 plus 3, judges it easy, and shortens its chain accordingly. It runs the deliberation it learned, and the deliberation it learned is long.

Here the natural assumption is that the extra tokens are at least earning their keep, making sure of an answer the model might otherwise fumble. They are not. In more than 92% of cases the first solution round already holds the correct answer. The rounds after it add little. By the fourth attempt, the share of genuinely new reasoning has fallen below 30%.¹ The model reaches the right answer early and then keeps restating it.

So the everyday behavior is an efficiency cost, not a wrong answer. The same study shows how much slack there is. A self-training fix cut the tokens used on the MATH500 benchmark from 2,407.9 to 1,330.7, a 48.6% reduction, and accuracy moved only from 92.8% to 93.2%.¹ You can remove almost half the thinking and the answer barely notices.

For this half of the behavior, the reinforcement-learning story holds up well. The chains grew because training rewarded correct answers and long chains delivered them, and the learned policy over-deliberates whenever nothing tells it to stop. That is the settled account of why a reasoning model burns compute on trivial input.

The built traps where longer reasoning loses accuracy

The second half is where the behavior turns genuinely costly, and where the claim is easiest to overstate. It is true that more reasoning can lower accuracy. It is not true in general.

Gema and colleagues, working with Anthropic, built tasks designed to punish extra reasoning. Some padded a counting question with irrelevant distractors or let a spurious feature correlate with the answer. Others stacked long chains of constraints for the model to track. On these, extending a model's reasoning drove its accuracy down. On one counting task whose answer was always two, Claude Opus 4 fell from near 100% to roughly 85-90% as its chain lengthened. DeepSeek R1 dropped harder, from about 70% to about 30% once five distractors were present. OpenAI's o3 held steadier, partly because it wrote shorter chains to begin with.³ The result reached a general audience under the headline "why thinking longer makes models dumber."⁴

But the same paper draws the boundary that headline erases. On standard arithmetic benchmarks, the ordinary word problems these models are usually measured on, accuracy is maintained as reasoning grows.³ The collapse is not a property of thinking longer. It appears when the task contains something a longer chain can grab onto: a distractor to follow, a spurious pattern to trust, a constraint to lose. The paper's own reading puts that drop on a different cause than the wasted compute of the last section. There, extended reasoning amplifies these flawed shortcuts. The extra thinking is not idle. It is actively steering the model wrong.

There is a third direction, and it runs the opposite way. On genuinely hard problems the same models underthink. Wang and colleagues found that on AIME, incorrect answers used 225% more tokens than correct ones and switched between distinct lines of reasoning 418% more often. More than 70% of those wrong responses contained a correct thought the model had reached and then abandoned.⁵ A separate group found the matching pattern. For a given question, correct solutions are on average shorter than incorrect ones. And letting a weaker model revise its own answer flips a right answer to a wrong one more often than it rescues one.⁶

So a long chain is not one failure with one fix. On easy inputs it is wasteful repetition. On built traps it is amplified error. On hard problems it is a model wandering away from an answer it already found.

Solid engineering under an unsettled cause

Work the behavior down to the layer where nothing beneath it would change the answer, and two things are solid. Test-time compute scales with the number of tokens a model generates, for the reason the earlier lesson gave. And reinforcement learning that rewarded correct answers lengthened the chains, because on hard problems long chains were what reached the answer.² Those are settled engineering.

What is not settled is how to frame the overthinking itself. Three accounts are live at once. It may be a training artifact, the length reward pulling chains longer than a task needs. It may be a missing difficulty estimate, the model having no gauge to size the problem and set its effort. Or it may be an inference-budget problem, the absence of any rule that allocates a fixed compute budget across a question. The papers do not agree. The overthinking work treats it as an efficiency artifact of a model with no difficulty gauge.¹ The inverse-scaling work puts the accuracy drops on reasoning that amplifies flawed heuristics, a cause separate from the length reward.³ The revisiting work frames it as weak sequential scaling and unreliable self-revision.⁶ None of them claims a single proven cause.

The fixes that exist are engineering responses to the symptom, not answers to that question. Budget forcing caps the chain by cutting it off, or extends it by appending the word "Wait" when the model tries to stop, which nudges it to recheck and sometimes correct a step. On AIME 2024 that lifted one model from 50% to 57%.⁷ Length-controlled training goes further: the user names a token budget in the prompt, making test-time compute something an operator allocates. A 1.5-billion-parameter model trained this way surpassed GPT-4o at equal reasoning lengths.⁸ Both give a dial for length. Neither settles why the model needs one.

That is the honest bottom of it. Every fix so far is a length control that a person sets, from a token budget in the prompt to a forced "Wait," because the model still cannot tell on its own that 2 plus 3 did not need 901 tokens.