How does post-training on traces improve performance without semantic reasoning?
This explores why training a model on step-by-step reasoning traces boosts its performance even when those traces aren't doing real logical work — what the gains are actually coming from if not the meaning of the steps.
This explores why training a model on step-by-step reasoning traces boosts its performance even when those traces aren't doing real logical work — what the gains are actually coming from if not the meaning of the steps. The most striking evidence is that the content of the traces barely matters: models trained on deliberately corrupted, irrelevant reasoning steps hold their accuracy and sometimes generalize *better* out of distribution Do reasoning traces need to be semantically correct?. The traces seem to work as computational scaffolding — a shape that makes the model spend tokens productively — rather than as a chain of true premises. The same picture shows up at inference: R1's intermediate tokens carry no special execution semantics, are generated the same way as any other output, and invalid traces routinely produce correct answers, so the trace correlates with the answer through learned formatting, not through causal reasoning Do reasoning traces actually cause correct answers?.
If the meaning isn't doing the work, what is? Several notes converge on *form*. Chain-of-thought turns out to be constrained imitation — models reproduce the familiar shape of reasoning from training rather than performing fresh inference Does chain-of-thought reasoning reveal genuine inference or pattern matching?, What makes chain-of-thought reasoning actually work?. Training format shapes the strategy a model adopts about 7.5× more than the actual domain, demo position alone can swing accuracy 20%, and logically invalid prompts work as well as valid ones What makes chain-of-thought reasoning actually work?. So post-training on traces is teaching a *protocol* — when to plan, when to backtrack, how to lay tokens down — not a body of correct deductions. That's why the gains are real but the semantics are optional.
A second mechanism is that the traces are selecting capability that's already there rather than installing new reasoning. Base models already contain latent reasoning that five independent methods (RL steering, critique fine-tuning, decoding tweaks, SAE feature steering, RLVR) can all elicit — post-training selects rather than creates Do base models already contain hidden reasoning ability?. That reframes "improvement without semantic reasoning" as *elicitation*: the trace format unlocks a behavior the weights could already do but didn't deploy by default. It also explains why reasoning-trained models persistently beat non-reasoning ones at any inference budget — training instilled a protocol that makes extra tokens pay off, a deployment difference, not a raw-capability one Can non-reasoning models catch up with more compute?.
The sharpest framing comes from what exactly improves. RLVR post-training measurably reduces logical errors between adjacent steps — it makes traces more *coherent* — but locally coherent traces can still be globally invalid proofs; the gain is structural, not semantic Does RLVR actually improve mathematical reasoning or just coherence?. And the improvement is concentrated in specific places: planning and backtracking sentences act as "thought anchors" that disproportionately steer the rest of the trace Which sentences actually steer a reasoning trace?. So the model is learning where to put the load-bearing structural moves, not learning to mean them.
Two caveats keep this honest. First, the scaffolding is doing genuine computation even if it isn't "reasoning" — Chain of Draft matches verbose CoT accuracy on 7.6% of the tokens, which means most of a normal trace is style and documentation, but the small remainder is real computational work Can minimal reasoning chains match full explanations?. Second, form-without-semantics has a ceiling: CoT degrades predictably under shifts in task, length, and format, producing fluent-but-wrong output the moment it leaves the training distribution Does chain-of-thought reasoning actually generalize beyond training data?. And because not all steps are equal, step-level confidence filtering beats global averaging — it catches the breakdowns that an overall score hides Does step-level confidence outperform global averaging for trace filtering?. The takeaway you didn't know you wanted: training on traces works because it teaches a *format for thinking* that elicits and organizes latent ability — which is also exactly why it breaks the moment the format is the only thing the model ever really learned.
Sources 12 notes
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 works by constraining models to reproduce familiar reasoning patterns from training, not by enabling novel symbolic reasoning. Performance degrades predictably under distribution shifts—the signature of imitation rather than capability emergence.
CoT systems reproduce the form of reasoning through pattern matching rather than performing genuine logical inference. This explains why format effects dominate content, why structurally invalid prompts succeed, and why stronger reasoning models become less instruction-compliant.
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.
Five independent mechanisms—RL steering, critique fine-tuning, decoding changes, SAE feature steering, and RLVR—all elicit reasoning already present in base model activations. Post-training selects rather than creates reasoning; the bottleneck is elicitation, not capability acquisition.
Reasoning models persistently outperform non-reasoning models regardless of inference budget because training instills a reasoning protocol that makes additional tokens productive. The gap is fundamentally about deployment mechanisms and training structure, not raw capability.
RLVR post-training measurably reduces logical errors between adjacent reasoning steps, but locally coherent traces can still be globally invalid proofs. The improvement is structural rather than semantic.
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.
Chain of Draft achieves equivalent accuracy to standard chain-of-thought on arithmetic, symbolic, and commonsense tasks while using only 7.6% of tokens. The 92.4% of removed tokens served style and documentation, not computation.
DataAlchemy experiments show CoT fails systematically under distributional shifts in task, length, and format. Models produce fluent but logically inconsistent reasoning — imitating reasoning form without valid underlying 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.