Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Towards a Mechanistic Understanding of Large Reasoning Models: A Survey of Training, Inference, and Failures

Yi Hu, Jiaqi Gu, Ruxin Wang, Zijun Yao, Hao Peng, Xiaobao Wu, Jianhui Chen, Muhan Zhang, Liangming Pan

TL;DR

The paper investigates how LRMs reason by dissecting training dynamics, internal reasoning mechanisms, and common failures. It synthesizes findings on the complementary roles of SFT and RL, the two-stage RL dynamics that enable reasoning, and the internal representations that underlie planning, reflection, and backtracking. It also analyzes failures such as hallucination, unfaithfulness, overthinking, and safety issues, linking them to training and inference dynamics and offering mechanistic explanations. The work aims to move beyond performance reporting toward a principled, predictive theory of reasoning in LRMs, with practical implications for interpretability, training design, and safer deployment.

Abstract

Reinforcement learning (RL) has catalyzed the emergence of Large Reasoning Models (LRMs) that have pushed reasoning capabilities to new heights. While their performance has garnered significant excitement, exploring the internal mechanisms driving these behaviors has become an equally critical research frontier. This paper provides a comprehensive survey of the mechanistic understanding of LRMs, organizing recent findings into three core dimensions: 1) training dynamics, 2) reasoning mechanisms, and 3) unintended behaviors. By synthesizing these insights, we aim to bridge the gap between black-box performance and mechanistic transparency. Finally, we discuss under-explored challenges to outline a roadmap for future mechanistic studies, including the need for applied interpretability, improved methodologies, and a unified theoretical framework.

Towards a Mechanistic Understanding of Large Reasoning Models: A Survey of Training, Inference, and Failures

TL;DR

The paper investigates how LRMs reason by dissecting training dynamics, internal reasoning mechanisms, and common failures. It synthesizes findings on the complementary roles of SFT and RL, the two-stage RL dynamics that enable reasoning, and the internal representations that underlie planning, reflection, and backtracking. It also analyzes failures such as hallucination, unfaithfulness, overthinking, and safety issues, linking them to training and inference dynamics and offering mechanistic explanations. The work aims to move beyond performance reporting toward a principled, predictive theory of reasoning in LRMs, with practical implications for interpretability, training design, and safer deployment.

Abstract

Reinforcement learning (RL) has catalyzed the emergence of Large Reasoning Models (LRMs) that have pushed reasoning capabilities to new heights. While their performance has garnered significant excitement, exploring the internal mechanisms driving these behaviors has become an equally critical research frontier. This paper provides a comprehensive survey of the mechanistic understanding of LRMs, organizing recent findings into three core dimensions: 1) training dynamics, 2) reasoning mechanisms, and 3) unintended behaviors. By synthesizing these insights, we aim to bridge the gap between black-box performance and mechanistic transparency. Finally, we discuss under-explored challenges to outline a roadmap for future mechanistic studies, including the need for applied interpretability, improved methodologies, and a unified theoretical framework.
Paper Structure (50 sections, 2 figures)

This paper contains 50 sections, 2 figures.

Table of Contents

  1. Introduction
  2. Contribution and Uniqueness.
  3. Understanding LRM Training
  4. Roles of SFT and RL in Post-Training
  5. SFT explores, RL compresses: do reasoning models have capability boundaries?
  6. SFT learns, RL repairs: how much SFT cold-start is appropriate?
  7. SFT memories, RL organizes: how can we make RL exhibit generalization?
  8. Understanding RL Training Dynamics
  9. Understanding two-stage training process: how does the Aha moment arise?
  10. What internal effects does RL have on the model?
  11. Exploitation v.s. exploration: how to make RL stable and effective?
  12. Understanding LRM Reasoning
  13. General Reasoning Structures
  14. Topological structures.
  15. Specific Reasoning Behaviors
  16. ...and 35 more sections

Figures (2)

  • Figure 1: Taxonomy of mechanistic research on LRMs. We organize existing studies into three core dimensions: reasoning-oriented training (Sec \ref{['sec:training']}), reasoning mechanisms (Sec \ref{['sec:model']}), and unintended behaviors (Sec \ref{['sec:special_behavior']}). Within each dimension, we synthesize recent findings based on the key research questions being investigated in the literature.
  • Figure 2: Taxonomy of our paper and representative works for each direction.