AI's Latent Reasoning Turns Out to Be More Legible Than Thought

The so-called latent reasoning models (LRMs) drew attention because they reason not in words but in a continuous hidden state. This is cheap at inference time and, in theory, lets the model explore several solution paths at once — but the convenience comes at the cost of transparency: there is no way to follow such a model's train of thought in the usual sense, since it produces no readable text. In work presented at the COLM 2026 conference, Connor Dilgren and Sarah Wiegreffe set out to test whether this reasoning really is as impenetrable as it seems.

Logic: the reasoning tokens are almost unnecessary

The authors studied two modern models — Coconut and CODI — on logical inference tasks such as PrOntoQA and ProsQA. It turned out that the hidden reasoning steps here are for the most part redundant: if you force the model to “stop thinking” early, it almost always produces the very same answer. In other words, the high accuracy on logical tasks is explained not by reasoning during inference but by peculiarities of the training data. This casts doubt on the role attributed to such tokens in earlier work and, at the same time, explains why LRMs do not always outperform models with explicit reasoning.

Math: ordinary solution steps are hidden inside

Things get more interesting where reasoning really is required — on mathematical tasks. The researchers projected the hidden states back into the ordinary vocabulary and checked whether the familiar step-by-step solution was concealed in them. It turned out that for correctly solved examples the reference reasoning chain could be recovered in 65–93% of cases. That is, the model apparently carries out in the background exactly the mathematical operations one would expect.

To make sure this was no coincidence, the authors changed the numbers in the problem statement and observed how the hidden states reacted — this confirmed which operations were being performed. Finally, they proposed a way to extract a verified reasoning chain in natural language without knowing the reference solution in advance. Such a chain could be found for most correct answers and only for a minority of incorrect ones.

From this comes a curious side conclusion: the very ability to decode a coherent line of reasoning serves as a sign that the answer is most likely correct. Contrary to a view widespread in the community, modern models with hidden reasoning encode processes that are quite interpretable.

Source: Lobsters