Why does reasoning accuracy degrade beyond a critical thinking token threshold?
This explores why models that 'think' longer don't just plateau but actually get worse past a certain point — and what's actually breaking inside the longer reasoning chain.
This explores why reasoning accuracy peaks and then *declines* once a model generates too many thinking tokens — not why it merely stops improving. The corpus is unusually direct here: pushing thinking from ~1,100 to ~16,000 tokens dropped benchmark accuracy from 87.3% to 70.3%, a non-monotonic curve where models overthink easy problems and underthink hard ones Does more thinking time always improve reasoning accuracy? When does thinking too much actually hurt reasoning?. The shape is an inverted U: accuracy rises to a peak at some intermediate length, then falls — and notably the optimal length is *shorter* for more capable models, which gravitate toward concise chains as RL training improves them Why does chain of thought accuracy eventually decline with length?.
The most interesting answer to *why* comes from looking at what extra tokens actually do. Beyond the peak, the additional reasoning doesn't add signal — it inflates output variance and injects self-revision errors, so the model second-guesses a correct answer into a wrong one When does thinking too much actually hurt reasoning?. This connects to a deeper finding: in untrained models, 'thinking mode' is actively counterproductive, inducing self-doubt that degrades performance, and only RL training flips that same mechanism into productive gap analysis Does extended thinking help or hurt model reasoning?. So the degradation isn't a quantity problem so much as a *quality-of-continuation* problem — extra tokens are an invitation to overwrite good intermediate work with noise.
Here's the turn the reader might not expect: the thinking tokens may not be doing the reasoning at all. Corrupted, semantically irrelevant traces train models nearly as well as correct ones, suggesting traces act as computational scaffolding rather than genuine inference Do reasoning traces need to be semantically correct?. Other work argues the trace is stylistic mimicry — invalid traces routinely yield correct answers, so the text correlates with the answer via learned formatting, not functional logic Do reasoning traces actually cause correct answers? Why does chain-of-thought reasoning fail in predictable ways?. If the visible chain is largely scaffolding, then piling on more of it is mostly adding ways to drift off the manifold the model was actually trained on.
But not all tokens are equal, which sharpens the picture further. The real work concentrates in a small minority: only ~20% of tokens are high-entropy 'forking points' that drive learning Do high-entropy tokens drive reasoning model improvements?, and specific reflection tokens like 'Wait' and 'Therefore' spike in mutual information with the correct answer Do reflection tokens carry more information about correct answers?. Past the threshold, you're generating more low-information filler around those rare pivots — diluting rather than deepening. And there's a capability cost layered on top: training models to always reason step-by-step narrows their cognition, so they overthink ill-posed questions instead of disengaging What critical thinking skills do reasoning models actually lose?.
The practical catch is that the threshold is invisible until you cross it — it shifts with model, task, and difficulty, with no reliable predictor, though difficulty estimators and runtime confidence signals are emerging as ways to detect it dynamically How can we predict the optimal thinking token threshold?. One provocative escape hatch: latent reasoning in continuous space scales test-time compute through hidden-state iteration without emitting verbalized tokens at all, hinting that the verbalization — the very thing that degrades past threshold — may be a training artifact rather than a requirement for thinking Can models reason without generating visible thinking tokens?.
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.
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.
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.
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.
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.
CoT guides models to pattern-match reasoning structure rather than perform genuine inference. This explains distribution-bounded failures, why structural coherence matters more than content correctness, and why performance optimizes against interpretability.
Only ~20% of tokens exhibit high entropy as pivotal reasoning decision points; RLVR primarily adjusts these forking tokens. Training exclusively on them matches or exceeds full-gradient performance, revealing that the minority carries the learning signal.
Specific tokens like "Wait" and "Therefore" show sharp spikes in mutual information with correct answers. Suppressing them harms reasoning while suppressing equal random tokens does not, and representation recycling improves accuracy 20%.
Models trained for step-by-step reasoning excel at in-distribution logical tasks but lose critical abilities: they overthink ill-posed questions instead of disengaging, and reason their way to wrong rules on inductive tasks. This cognitive narrowing is partly reversible through targeted RL training.
The overthinking threshold depends on task difficulty, model training, and domain, but remains invisible until crossed. Recent work suggests difficulty estimators and runtime confidence signals can detect thresholds dynamically.
Multiple architectures—depth-recurrent models, Heima, and Coconut—demonstrate that test-time compute scales through hidden state iteration rather than token generation. This suggests verbalization is a training artifact, not a reasoning requirement.