Study: Patterns in Latent Reasoning Models Don't Prove How They Reason

October 10, 2026, 14:54Source: NAVER LABS Europe

Patterns that researchers have read as signs of internal reasoning in latent reasoning models also show up in control models, according to a new paper from NAVER LABS Europe. The authors argue that such patterns are not explanations and that interpretability work needs matched controls and causal tests.

Key points

NAVER LABS Europe paper analyses two latent reasoning models, Coconut and CODI, against controls lacking the proposed recurrence or curriculum.
Patterns seen as evidence of internal reasoning, such as BFS-like frontiers and decodable arithmetic, also appear in the controls.
Causal interventions show latent-thought use is graded, scaling with a thought's causal effect on behavior.
The effect concentrates in low-rank directions whose geometry grows more structured as behavioral influence rises.
The authors say LRM interpretability requires matched controls and causal tests; the paper is due at EMNLP 2026.

What the paper examines

The paper, titled "Observable patterns are not explanations: a causal-geometric analysis of latent reasoning models," is by Darpan Aswal, Thomas Palmeira Ferraz, Yongxin Zhou and Maxime Peyrard. It is slated for the Conference on Empirical Methods in Natural Language Processing (EMNLP) in Budapest, Hungary, from 24 to 29 October, 2026, and was published on the NAVER LABS Europe site on 10 October 2026.

Latent reasoning models (LRMs) replace explicit chain-of-thought with continuous thoughts. Instead of writing out reasoning steps in text, they carry intermediate computation in continuous internal states. Recent work has treated observable patterns in these latent states as evidence for internal reasoning mechanisms. The paper names two examples: BFS-like frontiers and decodable arithmetic computation.

Controls challenge the evidence

The authors evaluated two LRMs, Coconut and CODI. They compared them against controls that lack the proposed recurrence or curriculum. According to the paper, the patterns previously taken as evidence also appear in those controls. They also do not always causally affect the model's behavior.

This matters because a pattern that appears with or without the supposed mechanism cannot point to that mechanism. The paper's title makes the point directly: what can be observed in a latent state is not the same as an explanation of how the model works.

Graded use of latent thoughts and low-rank geometry

Using causal interventions, the researchers found that the use of latent thoughts is not binary but graded. It scales with a thought's causal effect on model behavior.

Their geometric analyses show that this effect concentrates in low-rank directions. The step-to-step geometry of those directions becomes more structured as their behavioral influence increases.

What it means for interpretability

The authors conclude that latent thoughts should be treated as hidden computation, not hidden explanation. In their words, decodability, attention, or static structure alone cannot establish mechanism.

For researchers working on LRM interpretability, the practical message is methodological. The paper states that this work requires matched controls and causal tests. Claims that a decodable or structured latent pattern reveals how a model reasons would need to pass those checks before being accepted.