The LLM's reasoning would be latent, not in the chain of thought

A position paper by arXiv argues that the reasoning of language models should be studied as a trajectory of internal states, not through the chains of thought they display.

In a position paper posted on arXiv on April 17, 2026, Wenshuo Wang defends a thesis that is disturbing for the field of interpretability: what language models write when they "reason out loud" is not the place where reasoning takes place. This is played out in the formation of trajectories of latent states, the visible chain of thought being only a partial by-product.

The author formalizes the debate in three competing hypotheses: H1, the reasoning mainly passes through trajectories of latent states; H2, it operates mainly via the explicit chain of thought; and H0, the null hypothesis according to which the apparent progress is only due to an increase in serial calculation, without any specific representational mechanism.

Methodological consequences

By synthesizing recent empirical and mechanistic work, the article concludes that the available data mainly support H1, which it proposes to adopt as the default working hypothesis. The practical recommendation is to design experiments that separately measure three things that are currently confused: surface traces, latent states and the effect of the computational budget. If the thesis holds, a good part of the work basing interpretability on the reading of thought chains observes the wrong object.