Observable Patterns Are Not Explanations: A Causal-Geometric Analysis of Latent Reasoning Models

2026-06-10

AI Summary

This study addresses the challenge of distinguishing whether observable patterns in latent variable models arise from genuine reasoning mechanisms or superficial correlations. The work conceptualizes “latent thought” as an intrinsic computational process rather than a post hoc interpretive construct and introduces a systematic evaluation framework combining causal interventions, low-rank geometric analysis, control model design, and latent state decoding. Findings reveal that similar patterns emerge even in control models lacking reasoning capabilities, and that the latent variables genuinely influencing behavior exhibit gradient effects concentrated within a low-dimensional subspace. The research underscores the necessity of integrating causal testing with controlled experimentation in interpretability analyses and establishes a new paradigm for rigorously evaluating internal model mechanisms.

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 always causally 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-rank directions whose step-to-step 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 mechanism. LRM interpretability thus requires matched controls and causal tests.

Problem

• latent reasoning models
• causal analysis
• interpretability
• observable patterns
• hidden computation

Innovation

• latent reasoning models
• causal intervention
• geometric analysis
• low-rank structure
• model interpretability