Why do correct reasoning traces tend to be shorter than incorrect ones?
This explores why, across reasoning models, the chains that land on the right answer tend to use fewer tokens than the ones that fail — and what that says about whether longer thinking is actually better thinking.
This explores why correct reasoning traces tend to be shorter than incorrect ones — and the corpus suggests the answer is less flattering to "more thinking" than you'd expect. The most direct finding is that across o1-like models (QwQ, DeepSeek-R1, LIMO), correct solutions simply average fewer tokens, and the extra length in wrong answers isn't productive deliberation — it's self-revision that introduces and compounds errors rather than fixing them Why do correct reasoning traces contain fewer tokens?. A model that keeps second-guessing itself tends to talk itself out of a right answer, not into one.
That connects to a broader pattern: accuracy doesn't climb forever with length. Optimal chain-of-thought length follows an inverted-U — performance peaks at some intermediate length and then declines, and the more capable the model, the shorter its sweet spot. Reinforcement learning naturally pushes improving models toward shorter chains, meaning brevity is something competent models earn, not something forced on them Why does chain of thought accuracy eventually decline with length?. So short-and-correct and long-and-wrong are two faces of the same curve.
Why does length so often signal trouble? Part of the answer is that length isn't tracking difficulty the way we assume. In controlled maze experiments, trace length correlates with problem difficulty only on familiar in-distribution problems and decouples entirely once the problem is novel — length mostly reflects how close the task is to training schemas the model can recall, not how hard it's genuinely working Does longer reasoning actually mean harder problems?. When a model wanders into unfamiliar territory, it generates more tokens not because it's reasoning harder but because it's lost. The corpus names this failure directly: reasoning models "explore like tourists," wandering down invalid paths and abandoning promising ones prematurely, with the bloat coming from structural disorganization rather than insufficient compute Why do reasoning models abandon promising solution paths?.
Here's the part you might not expect to want to know: the extra tokens may not be "reasoning" at all in the functional sense. Several notes argue traces are stylistic mimicry — invalid or even deliberately corrupted traces produce correct answers nearly as often as valid ones, suggesting the trace is computational scaffolding and learned formatting, not a causal chain of logic Do reasoning traces actually cause correct answers? Do reasoning traces need to be semantically correct? What makes chain-of-thought reasoning actually work?. If that's true, then a longer trace gives more room for the formatting to drift, for a confident-but-wrong pivot to take hold, and for errors to accumulate — without a real logical engine pulling things back toward the right answer.
The practical upshot runs through the evaluation and filtering work: because length and verbosity are unreliable signals, you get more from watching *where* a trace turns than from how long it runs. Step-level confidence catches breakdowns that whole-trace averaging hides and lets you stop early before a trace bloats into failure Does step-level confidence outperform global averaging for trace filtering?, and the influential moments are sparse — a few planning and backtracking sentences act as "thought anchors" that steer everything after them Which sentences actually steer a reasoning trace?. In other words, correct traces aren't shorter because shortness is virtuous — they're shorter because getting it right means hitting the right anchor early and not needing the wandering, the revising, and the second-guessing that pad out the ones that miss.
Sources 9 notes
Across QwQ, DeepSeek-R1, and LIMO, correct solutions average fewer tokens than incorrect ones. Longer traces correlate with more self-revisions, which introduce and compound errors rather than improve reasoning 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.
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.
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.
R1's intermediate tokens carry no special execution semantics and are generated identically to other LLM output. Invalid traces frequently produce correct answers, proving traces are not causally necessary—they correlate with answers via learned formatting, not functional reasoning.
Models trained on systematically irrelevant traces maintain solution accuracy and sometimes improve out-of-distribution generalization, suggesting traces function as computational scaffolding rather than meaningful reasoning steps.
Research shows training format shapes reasoning strategy 7.5× more than domain, demo position swings accuracy 20%, and invalid CoT prompts work as well as valid ones. CoT is pattern-guided generation, not formal logic.
Local step-level confidence catches reasoning breakdowns that global averaging masks and enables early stopping before traces complete. This approach achieves comparable accuracy gains to naive majority voting with far fewer generated traces, proving trace quality matters more than quantity.
Counterfactual resampling, attention analysis, and causal suppression all identify planning and backtracking sentences as thought anchors—sparse critical points that guide subsequent reasoning. These are functional pivots, not noise.