Why doesn't letting an AI 'think longer' always make it more accurate — and can extra thinking actually backfire?
Why does increased reasoning effort fail to improve accuracy consistently?
This explores why giving a model more thinking time, through longer reasoning chains or more thinking tokens, sometimes helps, sometimes does nothing, and sometimes makes answers worse.
This explores why more thinking doesn't reliably produce better answers, and what is actually going on inside a model when it 'reasons harder.' The short version from the corpus is that accuracy rises with extra reasoning up to a point and then falls. One study found that raising thinking tokens from about 1,100 to 16,000 dropped benchmark accuracy from 87.3% to 70.3% Does more thinking time always improve reasoning accuracy?. The same pattern shows up across studies of test-time compute When does thinking too much actually hurt reasoning?. The best length isn't fixed, either. Harder problems need longer chains, but more capable models do best with shorter ones, and RL training tends to push models toward brevity on its own Why does chain of thought accuracy eventually decline with length?.
The less obvious part is how the extra reasoning hurts. It isn't just wasted effort. When researchers track the moments where a model changes its answer partway through a trace, they find models often reach the right answer and then talk themselves out of it Why does more reasoning sometimes make models worse?. A related finding changes what 'thinking longer' means. Longer traces seem to help mainly by widening the range of answers the model might give, which raises the odds that a correct one shows up. They don't seem to sharpen the reasoning itself. Past a certain length the spread gets too wide and accuracy collapses Does extended thinking actually improve reasoning or just increase variance?. On this view, extra thinking works more like rolling more dice than like thinking more carefully.
Training decides whether extra thinking helps at all. Untrained models in thinking mode tend to slide into unproductive self-doubt. After RL training, the same extended thinking turns into useful checking for gaps in the argument Does extended thinking help or hurt model reasoning?. So the question isn't only how much a model reasons but whether it has learned what to do with the extra room. It also matters where the gain is supposed to come from. Some work suggests factual knowledge sits in a model's lower layers and reasoning adjustments happen in higher ones. That would explain why reasoning training helps on math but can hurt knowledge-heavy fields like medicine, where more reasoning can't make up for a fact the model doesn't have Why does reasoning training help math but hurt medical tasks?.
Underneath all of this is an unsettling possibility: the visible reasoning may not be what produces the answer. Chain-of-thought examples with deliberately invalid logic improve performance almost as much as valid ones. That suggests models pick up the shape of reasoning rather than real inference Does logical validity actually drive chain-of-thought gains?. Supervised fine-tuning can raise final-answer accuracy while making each reasoning step about 39% less informative, so a correct answer can sit on top of after-the-fact rationalization Does supervised fine-tuning improve reasoning or just answers?. If the trace is partly decoration, there's no reason to expect a longer trace to be a better one.
One more angle to follow: length hurts on the input side too. Reasoning accuracy drops from 92% to 68% with just 3,000 tokens of irrelevant padding, far below the context limit, and chain-of-thought doesn't fix it Does reasoning ability actually degrade with longer inputs?. Taken together, the corpus suggests models struggle to stay on track over long stretches of text, whether that text is the problem they were given or their own reasoning. That makes knowing when to stop thinking a skill to train, not a free benefit of more compute.
Sources 10 notes
Increasing thinking tokens from ~1,100 to ~16K reduced benchmark accuracy from 87.3% to 70.3%, revealing a non-monotonic relationship where models overthink easy problems and underthink hard ones.
Empirical studies demonstrate non-monotonic scaling in test-time reasoning: accuracy peaks at a critical thinking-token count, then declines sharply (87.3% to 70.3% as tokens scale from 1,100 to 16,000). Extended thinking inflates output variance and introduces self-revision errors rather than improving solution quality.
Task accuracy peaks at intermediate CoT length, with optimal length increasing alongside task difficulty but decreasing with model capability. RL training naturally gravitates toward shorter chains as models improve, revealing that simplicity emerges from reward signals rather than explicit training.
Tracking flip events shows that extra reasoning tokens don't just hit diminishing returns—they actively cause models to second-guess and overwrite previously-correct answers, making accuracy non-monotonic in trace length.
Longer thinking traces improve accuracy through variance expansion—broader output distributions cover correct answers more often—not through better reasoning. Beyond a critical threshold, the distribution becomes too diffuse and accuracy drops, revealing the mechanism is sampling coverage, not genuine reasoning improvement.
Show all 10 sources
Vanilla models use thinking mode counterproductively, inducing self-doubt that degrades performance. RL training reverses this, transforming the same mechanism into beneficial gap analysis. Training mediates reasoning quality, not just quantity.
Two-phase inference model shows knowledge retrieval operates in lower network layers while reasoning adjustment happens in higher layers. This separation explains why reasoning training improves math but can degrade knowledge-intensive domains like medicine.
Illogical chain-of-thought exemplars matched valid CoT performance on BIG-Bench Hard, showing that structural properties—not logical validity—drive the gains. The model learns the form of reasoning, not genuine inference.
Supervised fine-tuning improves final-answer accuracy on benchmarks but cuts Information Gain by 38.9 percent, meaning models generate correct answers through post-hoc rationalization rather than genuine inferential steps. Standard metrics miss this degradation because they only measure final correctness.
FLenQA shows reasoning accuracy drops from 92% to 68% at just 3000 tokens of padding, far below context window capacity. The degradation is task-agnostic, uncorrelated with language modeling performance, and persists even with chain-of-thought prompting.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Think Deep, Not Just Long: Measuring LLM Reasoning Effort via Deep-Thinking Tokens
- Mining Hidden Thoughts from Texts: Evaluating Continual Pretraining with Synthetic Data for LLM Reasoning
- When More Thinking Hurts: Overthinking in LLM Test-Time Compute Scaling
- Does Thinking More always Help? Understanding Test-Time Scaling in Reasoning Models
- The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity
- Evaluating Theory of Mind in Reasoning Models: Robustness over Reasoning
- When More is Less: Understanding Chain-of-Thought Length in LLMs
- Between Underthinking and Overthinking: An Empirical Study of Reasoning Length and correctness in LLMs