Paper argues LLM reasoning is latent, and the chain of thought is not the reasoning
A new paper contends that latent-state dynamics should be the default object of study for LLM reasoning, and that evaluation designs must explicitly separate surface traces from latent states and from serial compute. If it holds, the visible chain of thought is a report about the reasoning rather than the reasoning itself.
The consequences run straight into practice. A large amount of current safety and evaluation work reads chains of thought as evidence of process — checking whether the stated steps are sound and treating that as a check on the computation. If the trace is a downstream artefact rather than the mechanism, then a correct-looking trace is weak evidence about what actually happened.
It also reframes what chain-of-thought prompting does. Under the surface reading, writing steps out is the model thinking. Under the latent reading, it is the model allocating more serial compute and conditioning on its own output, with the legibility a side effect. Both explanations predict better answers; only one predicts that the words are trustworthy.
The methodological demand is the durable contribution. Asking that experiments disentangle traces, latent states and serial compute is a design standard that can be applied to existing work, and a fair amount of it does not currently meet the bar.
arXiv — LLM reasoning is latent, not the chain of thought → · arXiv — Measuring weak-to-strong legibility of reasoning models → · arXiv — Thinking with reasoning skills: fewer tokens, more accuracy →