LLM Reasoning Is Latent, Not the Chain of Thought

2026-04-20 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, extended

Summary

A position paper by Wenshuo Wang from South China University of Technology argues that large language model (LLM) reasoning should be primarily understood as latent-state trajectory formation, rather than explicit surface chain-of-thought (CoT). The paper formalizes three competing hypotheses: H1 (latent-state trajectories mediate reasoning), H2 (explicit surface CoT mediates reasoning), and H0 (generic serial compute explains reasoning gains). Through a reorganization of existing empirical and mechanistic work, and new compute-audited exemplars, the author finds that current evidence most strongly supports H1 as a default working hypothesis, though not as a task-independent verdict. The paper recommends treating latent-state dynamics as the default object of study for LLM reasoning and evaluating reasoning with designs that explicitly disentangle surface traces, latent states, and serial compute.

Key takeaway

For research scientists investigating LLM reasoning, you should shift your default perspective to latent-state dynamics rather than relying solely on surface chain-of-thought. This means designing experiments that explicitly separate and control for surface traces, latent states, and generic serial compute to accurately determine causal leverage and improve interpretability and safety monitoring.

Key insights

LLM reasoning is best understood as latent-state trajectory formation, not explicit chain-of-thought.

Principles

• Latent-state dynamics are the default object of LLM reasoning study.
• Reasoning evaluation must disentangle surface traces, latent states, and serial compute.

Method

A diagnostic experimental design for LLM reasoning must factorize surface traces ($S$), latent-state trajectories ($Z$), and generic serial compute ($B$), use matched controls, and define differential verdict rules.

In practice

• Use compute-audited designs to compare reasoning hypotheses.
• Prioritize latent-state dynamics in LLM reasoning research.

Topics

• LLM Reasoning
• Latent State Dynamics
• Chain of Thought
• Serial Compute
• Reasoning Regimes

Best for: Research Scientist, AI Scientist, Machine Learning Engineer

Related on AIssential

• LLM Reasoning Is Latent, Not the Chain of Thought — LLM reasoning is best understood as latent-state trajectory formation, not explicit chain-of-thought.
• Why We Think — Allocating more test-time compute via methods like Chain-of-Thought significantly boosts LLM reasoning and problem-solving capabilities.
• Reasoning as Pattern Matching: Shared Mechanisms in Human and LLM Everyday Reasoning — Everyday causal reasoning in humans and LLMs appears driven by pattern-matching, not abstract world models.
• The Sequence AI of the Week #867: Thinking in Latents: Why Sapient's HRM-Text Is a Quiet Rebuke to Chain-of-Thought — Chain-of-Thought is an inefficient workaround for Transformer's inherent shallow computational depth.
• How Do Answer Tokens Read Reasoning Traces? Self-Reading Patterns in Thinking LLMs for Quantitative Reasoning — Benign self-reading patterns in LLMs, marked by focused, forward-shifting attention, correlate with correct quantitative reasoning.
• Correct Chains, Wrong Answers: Dissociating Reasoning from Output in LLM Logic — LLMs can produce correct reasoning steps but still output wrong final answers, revealing a reasoning-output dissociation.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.