Do Latent-CoT Models Think Step-by-Step? A Mechanistic Study on Sequential Reasoning Tasks
cs.AI, cs.LG
Submitted: 2026-01-31
Updated: 2026-09-27
Comments: 32 pages, 20 figures
Code: https://github.com/TransformerLensOrg/TransformerLens
License: http://creativecommons.org/licenses/by/4.0/
The gist: Latent Chain-of-Thought aims to enable step-by-step computation without emitting long rationales, yet its internal mechanisms remain unclear.
Terminology
Abstract
Latent Chain-of-Thought aims to enable step-by-step computation without emitting long rationales, yet its internal mechanisms remain unclear. We study CODI, a continuous-thought teacher-student distillation model, on strictly sequential polynomial-iteration tasks with known intermediate states. Using logit-lens decoding, linear probes, attention analysis, and activation patching, we localize intermediate-state representations and trace how they are routed to the final readout. In short-horizon, low-hop tasks, CODI forms faithful bridge states across latent-thought positions, while the final input follows a separate near-direct route; predictions arise through late fusion at the answer readout. As task depth and difficulty increase, however, CODI does not reliably sustain a full latent rollout: it either compresses computation into a partial late-intermediate pathway or, in harder regimes, loses the latent reasoning signature altogether. To explain this transition, we show theoretically that the task's algebraic structure controls its effective memory: compressible regimes support late-bottleneck reasoning, while incompressible regimes preserve full-history dependence and destabilize latent rollouts. Overall, our results characterize when CODI-style latent-CoT yields faithful iterative computation versus compressed or shortcut strategies.
Sources
- Understanding intermediate layers using linear classifier probes
- Hopping Too Late: Exploring the Limitations of Large Language Models on Multi-Hop Queries
- A Mechanistic Analysis of a Transformer Trained on a Symbolic Multi-Step Reasoning Task
- From Explicit CoT to Implicit CoT: Learning to Internalize CoT Step by Step
- How to think step-by-step: A mechanistic understanding of chain-of-thought reasoning
- DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning
- How Do LLMs Perform Two-Hop Reasoning in Context?
- Training Large Language Models to Reason in a Continuous Latent Space
- Towards a Mechanistic Interpretation of Multi-Step Reasoning Capabilities of Language Models
- OpenAI o1 System Card
- Investigating Multi-Hop Factual Shortcuts in Knowledge Editing of Large Language Models
- Adam: A Method for Stochastic Optimization
- LLMs Faithfully and Iteratively Compute Answers During CoT: A Systematic Analysis With Multi-step Arithmetics
- Understanding and Patching Compositional Reasoning in LLMs
- The Expressive Power of Transformers with Chain of Thought
- An Investigation of Neuron Activation as a Unified Lens to Explain Chain-of-Thought Eliciting Arithmetic Reasoning of LLMs
- Reasoning with Latent Thoughts: On the Power of Looped Transformers
- CODI: Compressing Chain-of-Thought into Continuous Space via Self-Distillation
- To CoT or not to CoT? Chain-of-thought helps mainly on math and symbolic reasoning
- Grokked Transformers are Implicit Reasoners: A Mechanistic Journey to the Edge of Generalization
Related papers
- MAVEN-T: Reinforced Heterogeneous Distillation for Real-Time Multi-Agent Trajectory Prediction
- Model Discovery Agent: LLM-assisted Bayesian experiment design for data-efficient discovery of mechanistic world models
- Agents' Last Exam
- The Clinician's Veto: Navigating Trust, Liability, and Uncertainty in Autonomous AI Prescribing
- MindHelper: Closed-Loop Embodied Mental-State Reasoning for Precision Intervention
- Incumbent Advantage: Brand Bias and Cognitive Manipulation Dynamics in LLM Recommendation Systems