Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Opening the Black Box: A Survey on the Mechanisms of Multi-Step Reasoning in Large Language Models

Liangming Pan, Jason Liang, Jiaran Ye, Minglai Yang, Xinyuan Lu, Fengbin Zhu

TL;DR

This survey addresses how large language models perform multi-step reasoning by examining two paradigms: implicit reasoning hidden in internal activations and explicit Chain-of-Thought (CoT) reasoning that externalizes steps. It organizes the literature around seven interconnected research questions, spanning mechanisms in hidden states, training dynamics, and how verbalized reasoning reshapes computation, along with five future directions. The authors synthesize evidence on layer specialization, depth bottlenecks, and shortcuts that can undermine genuine reasoning, as well as mechanisms by which CoT expands effective depth, modularity, and robustness, while noting CoT’s potential unfaithfulness as an explanation. The work aims to ground advances in mechanistic understanding to enable causal interventions, faithful evaluation, and safer, more controllable reasoning in transformer-based LLMs.

Abstract

Large Language Models (LLMs) have demonstrated remarkable abilities to solve problems requiring multiple reasoning steps, yet the internal mechanisms enabling such capabilities remain elusive. Unlike existing surveys that primarily focus on engineering methods to enhance performance, this survey provides a comprehensive overview of the mechanisms underlying LLM multi-step reasoning. We organize the survey around a conceptual framework comprising seven interconnected research questions, from how LLMs execute implicit multi-hop reasoning within hidden activations to how verbalized explicit reasoning remodels the internal computation. Finally, we highlight five research directions for future mechanistic studies.

Opening the Black Box: A Survey on the Mechanisms of Multi-Step Reasoning in Large Language Models

TL;DR

This survey addresses how large language models perform multi-step reasoning by examining two paradigms: implicit reasoning hidden in internal activations and explicit Chain-of-Thought (CoT) reasoning that externalizes steps. It organizes the literature around seven interconnected research questions, spanning mechanisms in hidden states, training dynamics, and how verbalized reasoning reshapes computation, along with five future directions. The authors synthesize evidence on layer specialization, depth bottlenecks, and shortcuts that can undermine genuine reasoning, as well as mechanisms by which CoT expands effective depth, modularity, and robustness, while noting CoT’s potential unfaithfulness as an explanation. The work aims to ground advances in mechanistic understanding to enable causal interventions, faithful evaluation, and safer, more controllable reasoning in transformer-based LLMs.

Abstract

Large Language Models (LLMs) have demonstrated remarkable abilities to solve problems requiring multiple reasoning steps, yet the internal mechanisms enabling such capabilities remain elusive. Unlike existing surveys that primarily focus on engineering methods to enhance performance, this survey provides a comprehensive overview of the mechanisms underlying LLM multi-step reasoning. We organize the survey around a conceptual framework comprising seven interconnected research questions, from how LLMs execute implicit multi-hop reasoning within hidden activations to how verbalized explicit reasoning remodels the internal computation. Finally, we highlight five research directions for future mechanistic studies.
Paper Structure (37 sections, 1 figure)

This paper contains 37 sections, 1 figure.

Table of Contents

  1. Introduction
  2. Implicit Multi-Step Reasoning
  3. What are the internal mechanisms of latent multi-step reasoning?
  4. Functional specialization of layers.
  5. Uncovering fine-grained reasoning structures.
  6. Layer depth as the primary bottleneck for implicit reasoning.
  7. Why implicit reasoning sometimes fails.
  8. How latent multi-step reasoning capability is acquired during training?
  9. Grokking marks the shift from memorization to reasoning.
  10. Factors influencing the emergence of reasoning.
  11. To what extent does multi-step reasoning rely on shortcuts?
  12. Factual shortcuts bypass intermediate reasoning.
  13. Shortcuts based on surface-level pattern matching.
  14. Explicit Multi-Step Reasoning
  15. Where and When Does CoT Help?
  16. ...and 22 more sections

Figures (1)

  • Figure 1: The cognitive framework and organizational structure of this survey. We explore the mechanisms of multi-step reasoning through two distinct paradigms: Implicit Reasoning and Explicit Reasoning, through seven interconnected Research Questions. The bottom panel highlights five strategic directions for future research.