If reasoning is latent, the chain of thought is a receipt

A receipt tells you a transaction happened. It does not prove the transaction was the one described. A new paper argues the visible reasoning trace stands in exactly that relationship to the computation that produced it.

A new paper argues latent-state dynamics should be the default object of study for LLM reasoning, and that evaluation must explicitly separate surface traces, latent states and serial compute. If that holds, a great deal of current practice is reading the wrong artefact.

What breaks

Substantial safety and evaluation work reads chains of thought as evidence of process — check whether the stated steps are sound, treat that as a check on the computation. If the trace is a downstream report rather than the mechanism, a correct-looking trace is weak evidence about what actually happened, and a model that produces sound-looking steps for unsound reasons is invisible to the method.

It also reframes chain-of-thought prompting. Under the surface reading, writing steps out is the model thinking. Under the latent reading, it is the model buying more serial compute and conditioning on its own output, with legibility a side effect. Both predict better answers. Only one predicts the words can be trusted.

The second finding makes it worse

Reasoning appears to collapse rather than degrade as problems get harder. Smooth degradation means a score at one difficulty tells you something about performance just past it. A sharp transition means it does not — 95% just below the boundary can be near-useless just above, and testing on the easy side never reveals where the edge is.

Anyone sizing a model to a task by benchmarking representative examples is implicitly assuming continuity. If the transition is sharp and the task distribution has a tail, failures cluster in exactly the hard cases where being wrong costs most. That is the worst possible correlation between error rate and stakes.

Fewer tokens, more accuracy

The related result that reasoning skills produce fewer tokens and higher accuracy cuts against a deeply embedded intuition, and it points the same direction as the latent-reasoning argument: length of visible deliberation is not a proxy for quality of reasoning. Products have been designed around that proxy. So have evaluations.

The durable contribution

The methodological demand outlasts any particular claim. Asking that experiments disentangle traces, latent states and serial compute is a design standard applicable to existing work — and a fair amount of that work does not currently meet it. Standards like this are how a field stops accumulating results that cannot be compared.

arXiv — LLM reasoning is latent, not the chain of thought → · arXiv — Logical phase transitions: understanding collapse in LLM logical reasoning → · arXiv — Thinking with reasoning skills: fewer tokens, more accuracy →