Giving AI more time to think backfires on easy problems — in one study, extra thinking cut accuracy from 87% to 70%.
When does extended thinking hurt performance on easier problems?
This explores the specific conditions under which making a model 'think longer' backfires on problems that don't need it — and what the corpus says is actually going wrong when it does.
This explores when extended thinking hurts easy problems specifically — and the corpus points to a clear pattern: the damage is worst precisely where the problem was already within reach. The headline number recurs across several notes: accuracy falls from 87.3% to 70.3% as thinking tokens scale from ~1,100 to 16,000, a non-monotonic curve where accuracy peaks early then declines Does more thinking time always improve reasoning accuracy?, Does more thinking time actually improve LLM reasoning?, When does thinking too much actually hurt reasoning?. The 'more thinking is always better' assumption is one of the most falsifiable claims in the field — and it fails.
The most concrete answer to *when* comes from the flip-event work: extra reasoning tokens don't just hit diminishing returns, they actively cause models to second-guess and overwrite answers they had already gotten right Why does more reasoning sometimes make models worse?. On an easy problem the model often reaches the correct answer early, so every additional token is downside risk — more chances to talk itself out of a good answer. That reframes overthinking from wasted compute into an active hazard.
Why does the extra length not help? One note argues the mechanism was never 'better reasoning' in the first place: longer traces raise accuracy by inflating output *variance* — a broader distribution simply covers the right answer more often — and past a critical threshold the distribution gets too diffuse and accuracy drops Does extended thinking actually improve reasoning or just increase variance?. Easy problems have a narrow correct target, so widening the distribution helps little and eventually hurts. This connects to a deeper finding that trace length tracks how close a problem sits to the training distribution, not how hard it genuinely is Does longer reasoning actually mean harder problems? — the model isn't allocating compute to difficulty, it's recalling schemas.
The cleaner framing is that optimal thinking length follows an inverted-U that shifts with both difficulty and model capability: shorter for easy tasks, longer for hard ones, and *shorter overall as models get more capable* Why does chain of thought accuracy eventually decline with length?. Fixed or uniform token budgets are the real culprit — they over-think the easy and under-think the hard Does thinking time need to scale with problem difficulty?. So 'when does it hurt easy problems' has a precise answer: whenever the budget exceeds what that problem needed, which for easy problems is almost immediately.
Here's the twist worth carrying away: overthinking isn't inherent to thinking — it's a training artifact. Vanilla models use thinking mode counterproductively, spiraling into self-doubt, while RL training redirects the very same mechanism into productive gap-analysis Does extended thinking help or hurt model reasoning?, Why does asking models to think first hurt performance?. Related failures — reasoning models that can't reject ill-posed questions and pad them with redundant steps Why do reasoning models overthink ill-posed questions?, or that wander and abandon good paths like tourists rather than scientists Why do reasoning models abandon promising solution paths? — suggest the deeper problem isn't too much compute but never learning *when to stop*.
Sources 12 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.
Accuracy drops from 87.3% to 70.3% as thinking tokens scale from 1,100 to 16,000, and bypassing explicit reasoning entirely matches or beats standard thinking at equal token budgets. The relationship is non-monotonic, not the linear improvement commonly assumed.
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.
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 12 sources
Controlled A* maze experiments show trace length correlates with difficulty only in-distribution but decouples entirely out-of-distribution. Trace length primarily reflects recall of training schemas, not adaptive computation.
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.
Research shows the token budget that maximizes accuracy differs by problem difficulty. Fixed budgets either over-think easy problems or under-think hard ones, meaning adaptive allocation is a more efficient accuracy lever than increasing tokens uniformly.
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.
Prompting models to think before responding degrades performance on general tasks. RL training with judges evaluating only responses teaches models to generate thoughts that actually improve outputs across diverse task types, not just math.
Reasoning models generate redundant, lengthy responses to questions with missing premises while non-reasoning models correctly identify them as unanswerable. Training optimizes for producing reasoning steps but never teaches models when to disengage.
Reasoning LLMs exhibit two reinforcing failures: wandering (invalid exploration) and underthinking (premature path-switching). Decoding-level interventions like thought-switching penalties improve accuracy without fine-tuning, suggesting viable solutions exist but are abandoned prematurely.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Mining Hidden Thoughts from Texts: Evaluating Continual Pretraining with Synthetic Data for LLM Reasoning
- The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity
- Think Deep, Not Just Long: Measuring LLM Reasoning Effort via Deep-Thinking Tokens
- Does Thinking More always Help? Understanding Test-Time Scaling in Reasoning Models
- When More Thinking Hurts: Overthinking in LLM Test-Time Compute Scaling
- Rethinking Thinking Tokens: LLMs as Improvement Operators
- Between Underthinking and Overthinking: An Empirical Study of Reasoning Length and correctness in LLMs
- Beyond Passive Critical Thinking: Fostering Proactive Questioning to Enhance Human-AI Collaboration