If the chain of thought is not the reasoning, three years of tooling was aimed at the wrong object

One paper argued that reasoning happens in latent space and the emitted chain is a rendering of it. The research agenda reorganised around that claim within a quarter — which is impressive, and fast enough to deserve some scrutiny.

The claim is that the process driving a model's answer is latent, and the chain of thought is not that process. If true, improving reasoning by improving that text and monitoring reasoning by reading it are both aimed at a shadow.

Why it landed so hard

Because it explains things practitioners already noticed and could not account for: models reaching correct answers through visibly incoherent chains, chains that change without the answer changing, and the persistent failure of chain-quality metrics to predict task success. A result that retrospectively explains a pile of anomalies gets adopted quickly.

The follow-on work is visible in the literature already — token-wise latent-explicit routing, adaptive information control for search-augmented reasoning, and a general move to treat the latent trajectory as the optimisation target.

The scrutiny it has not had

It is a strong claim and a convenient one. Convenient because it licenses discarding a monitoring surface that constrains capability work, and because it reframes an inconvenient artifact as epiphenomenal. Neither of those makes it wrong. Both are reasons to want replication before a field builds on it, and influential results in this area have not always received that.

How evaluation is moving in parallel

Static benchmarks leak into training data and can be optimised against without any underlying improvement. Opponent-based evaluation under incomplete information has no fixed answer key to memorise, because the right move depends on the other model. Fog of war makes it a reliability test rather than a reasoning test — which is the failure mode that shows up when agents run unsupervised.

Both moves point the same way: away from what a model emits, toward what it does. That is the right direction, and it is being taken faster than the verification tooling can follow.

arXiv — LLM Reasoning Is Latent, Not the Chain of Thought → · Sebastian Raschka — LLM Research Papers: The 2026 List →