Observable Patterns Are Not Explanations: A Causal-Geometric Analysis of Latent Reasoning Models
Abstract
Latent reasoning models (LRMs) replace explicit chain-of-thought with continuous thoughts. Recent work treats observable latent-state patterns, such as BFS-like frontiers and decodable arithmetic computation, as evidence for internal reasoning mechanisms. Evaluating two LRMs (Coconut and CODI) against controls lacking the proposed recurrence or curriculum, we find these patterns also appear in the controls and do not consistently affect behavior. Causal interventions reveal that latent-thought utilization is not binary but graded, scaling with a thought's causal effect on model behavior. Geometric analyses reveal this effect concentrates in low-dimensional subspaces whose step-wise geometry grows more structured as their behavioral influence increases. Latent thoughts should therefore be treated as hidden computation, not hidden explanation: decodability, attention, or static structure alone cannot establish LRM reasoning mechanisms; that requires matched controls and causal tests.
Related
- A Van Hiele–Newman model analysis of geometric reasoning and problem-solving errors among final-year students in a selected senior high school
- A Geometric Derivation of the Einstein Equations from the Causal Action Principle
- Explanations, Prompts, and Formalizations: Arguments for New Norms in LLM-Enabled Mathematical Research
- When Are Concept Bottleneck Model Explanations Faithful and Compact?
- Decision-Tree Surrogates for Weak Counterfactual Explanations in Integer Optimization
BibTeX
@misc{aswal2026observable,
title = {{Observable Patterns Are Not Explanations: A Causal-Geometric Analysis of Latent Reasoning Models}},
author = {Darpan Aswal and Thomas Palmeira Ferraz and Yongxin Zhou and Maxime Peyrard},
howpublished = {HAL},
year = {2026},
month = sep,
doi = {10.48550/arxiv.2606.12689},
eprint = {2606.12689},
archivePrefix = {arXiv},
url = {https://univ-grenoble-alpes.hal.science/hal-05661630v2},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Observable Patterns Are Not Explanations: A Causal-Geometric Analysis of Latent Reasoning Models
AU - Darpan Aswal
AU - Thomas Palmeira Ferraz
AU - Yongxin Zhou
AU - Maxime Peyrard
PY - 2026
DA - 2026/09/06
PB - HAL
DO - 10.48550/arxiv.2606.12689
UR - https://univ-grenoble-alpes.hal.science/hal-05661630v2
AB - Latent reasoning models (LRMs) replace explicit chain-of-thought with continuous thoughts. Recent work treats observable latent-state patterns, such as BFS-like frontiers and decodable arithmetic computation, as evidence for internal reasoning mechanisms. Evaluating two LRMs (Coconut and CODI) against controls lacking the proposed recurrence or curriculum, we find these patterns also appear in the controls and do not consistently affect behavior. Causal interventions reveal that latent-thought utilization is not binary but graded, scaling with a thought's causal effect on model behavior. Geometric analyses reveal this effect concentrates in low-dimensional subspaces whose step-wise geometry grows more structured as their behavioral influence increases. Latent thoughts should therefore be treated as hidden computation, not hidden explanation: decodability, attention, or static structure alone cannot establish LRM reasoning mechanisms; that requires matched controls and causal tests.
ER -