🚀 News

• [2026.1.29] Our survey is now available at: https://arxiv.org/abs/2601.19928
• [2026.1.11] 🔥 We are excited to release our survey on LRM mechanistic studies in this repo.

📌 Overview

Our paper provides a comprehensive survey of the mechanistic understanding of LRMs.

We organize recent findings into three core dimensions:

1. Training Dynamics:
   1. Roles of SFT and RL in Post-Training
   2. Understanding RL Training Dynamics
2. Reasoning Mechanisms:
   1. General Reasoning Structures
   2. Specific Reasoning Behaviors
   3. Internal Mechanisms
3. Unintended Behaviors:
   1. Hallucination
   2. Unfaithfulness
   3. Overthinking
   4. Unsafety

🔗 Citation

If you find our survey helpful, please consider citing our paper:

@misc{hu2026mechanisticunderstandinglargereasoning,
      title={Towards a Mechanistic Understanding of Large Reasoning Models: A Survey of Training, Inference, and Failures}, 
      author={Yi Hu and Jiaqi Gu and Ruxin Wang and Zijun Yao and Hao Peng and Xiaobao Wu and Jianhui Chen and Muhan Zhang and Liangming Pan},
      year={2026},
      eprint={2601.19928},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2601.19928}, 
}

🗂️ Contents

• Towards a Mechanistic Understanding of Large Reasoning Models: A Survey of Training, Inference, and Failures
  • 🚀 News
  • 📌 Overview
  • 🔗 Citation
  • 🗂️ Contents
  • 📄 Paper List
    • Training Dynamics
      • Roles of SFT and RL in Post-Training
      • Understanding RL Training Dynamics
    • Reasoning Mechanisms
      • General Reasoning Structures
      • Specific Reasoning Behaviors
      • Internal Mechanisms
    • Unintended Behaviors
      • Hallucination
      • Unfaithfulness
      • Overthinking
      • Unsafety
    • Training Methods
      • Combine SFT with RL
      • RL balancing exploration and exploitation

📄 Paper List

Training Dynamics

Roles of SFT and RL in Post-Training

Click to hide/show the paper list

Date	Name	Title	Paper	Github
2025-01	Deepseek-R1	DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning
2025-04	limit of RLVR	Does Reinforcement Learning Really Incentivize Reasoning Capacity in LLMs Beyond the Base Model?
2025-09	RL Squeezes, SFT Expands	RL Squeezes, SFT Expands: A Comparative Study of Reasoning LLMs		-
2025-07	Invisible Leash	The Invisible Leash: Why RLVR May or May Not Escape Its Origin		-
2025-09	Representation Geometry of LLM	Tracing the Representation Geometry of Language Models from Pretraining to Post-training		-
2025-05	Entropy Mechanism of RL	The Entropy Mechanism of Reinforcement Learning for Reasoning Language Models
2025-09	Emergent Attention Heads	Thinking Sparks!: Emergent Attention Heads in Reasoning Models During Post Training		-
2025-01	SFT Memorizes, RL Generalizes	SFT Memorizes, RL Generalizes: A Comparative Study of Foundation Model Post-training
2025-09	CoNet	How LLMs Learn to Reason: A Complex Network Perspective		-
2025-08	RL Is Neither a Panacea Nor a Mirage	RL Is Neither a Panacea Nor a Mirage: Understanding Supervised vs. Reinforcement Learning Fine-Tuning for LLMs		-
2025-09	RL Heals OOD Forgetting	RL Fine-Tuning Heals OOD Forgetting in SFT
2025-05	ProRL	ProRL: Prolonged Reinforcement Learning Expands Reasoning Boundaries in Large Language Models
2025-09	From f(x) and g(x) to f(g(x))	From f(x) and g(x) to f(g(x)): LLMs Learn New Skills in RL by Composing Old Ones		-
2025-12	From Atomic to Composite	From Atomic to Composite: Reinforcement Learning Enables Generalization in Complementary Reasoning
2025-06	OctoThinker	OctoThinker: Mid-training Incentivizes Reinforcement Learning Scaling
2025-12	Interplay LM Reasoning	On the Interplay of Pre-Training, Mid-Training, and RL on Reasoning Language Models

Understanding RL Training Dynamics

Click to hide/show the paper list

Date	Name	Title	Paper	Github
2025-09	HICRA	Emergent Hierarchical Reasoning in LLMs through Reinforcement Learning
2025-09	CoNet	How LLMs Learn to Reason: A Complex Network Perspective		-
2025-10	Two-Stage Dynamic	The Debate on RLVR Reasoning Capability Boundary: Shrinkage, Expansion, or Both? A Two-Stage Dynamic View		-
2025-09	RL Enhances Activation	Reinforcement Learning Fine-Tuning Enhances Activation Intensity and Diversity in the Internal Circuitry of LLMs
2025-09	Post-Training Structural Changes	Understanding Post-Training Structural Changes in Large Language Models		-
2025-10	AlphaRL	On Predictability of Reinforcement Learning Dynamics for Large Language Models
2025-05	Entropy Mechanism of RL	The Entropy Mechanism of Reinforcement Learning for Reasoning Language Models
2025-10	Reasoning Boundary Paradox	The Reasoning Boundary Paradox: How Reinforcement Learning Constrains Language Models
2025-09	VERL	Beyond the Exploration-Exploitation Trade-off: A Hidden State Approach for LLM Reasoning in RLVR

Reasoning Mechanisms

General Reasoning Structures

Click to hide/show the paper list

Date	Name	Title	Paper	Github
2025-04	DeepSeek-R1 Thoughtology	DeepSeek-R1 Thoughtology: Let's think about LLM Reasoning		-
2025-06	Functional Blocks	Beyond Accuracy: Dissecting Mathematical Reasoning for LLMs Under Reinforcement Learning
2025-06	Thought Anchors	Thought Anchors: Which LLM Reasoning Steps Matter?
2025-09	Schoenfeld's Episode Theory	Understanding the Thinking Process of Reasoning Models: A Perspective from Schoenfeld's Episode Theory
2025-11	ReJump	ReJump: A Tree-Jump Representation for Analyzing and Improving LLM Reasoning
2025-05	Structural Patterns	What Makes a Good Reasoning Chain? Uncovering Structural Patterns in Long Chain-of-Thought Reasoning		-
2025-06	Topology of Reasoning	Topology of Reasoning: Understanding Large Reasoning Models through Reasoning Graph Properties
2025-05	Mapping the Minds of LLMs	Mapping the Minds of LLMs: A Graph-Based Analysis of Reasoning LLM

Specific Reasoning Behaviors

Click to hide/show the paper list

Date	Name	Title	Paper	Github
2025-06	Thought Anchors	Thought Anchors: Which LLM Reasoning Steps Matter?
2025-05	Thought Anchors	Cognitive Behaviors that Enable Self-Improving Reasoners, or, Four Habits of Highly Effective STaRs
2025-04	Understanding Aha Moments	Understanding Aha Moments: from External Observations to Internal Mechanisms		-
2025-02	May Not Aha Moment	There may not be aha moment in r1-zero-like training - a pilot study
2025-10	First Try Matters	First Try Matters: Revisiting the Role of Reflection in Reasoning Models
2025-06	The Illusion of Thinking	The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity		-
2025-9	Performative Thinking	Performative Thinking? The Brittle Correlation Between CoT Length and Problem Complexity		-

Internal Mechanisms

Click to hide/show the paper list

Date	Name	Title	Paper	Github
2025-05	Towards Understanding Distilled Reasoning Models	Towards Understanding Distilled Reasoning Models: A Representational Approach
2025-05	Sparse Autoencoders	I Have Covered All the Bases Here: Interpreting Reasoning Features in Large Language Models via Sparse Autoencoders
2025-04	Hood of a Reasoning Model	Under the Hood of a Reasoning Model		-
2025-06	Steering Vectors	Understanding Reasoning in Thinking Language Models via Steering Vectors
2025-10	Thinking Models Learn When To Reason	Base Models Know How to Reason, Thinking Models Learn When
2025-06	Thought Anchors	Thought Anchors: Which LLM Reasoning Steps Matter?
2025-09	From Reasoning to Answer	From Reasoning to Answer: Empirical, Attention-Based and Mechanistic Insights into Distilled DeepSeek R1 Models
2025-04	Understanding Aha Moments	Understanding Aha Moments: from External Observations to Internal Mechanisms		-
2025-12	ReflCtrl	ReflCtrl: Controlling LLM Reflection via Representation Engineering
2025-08	Latent Directions of Reflection	Unveiling the Latent Directions of Reflection in Large Language Models
2025-07	Reasoning-Finetuning Repurposes Latent Representations	Reasoning-Finetuning Repurposes Latent Representations in Base Models
2025-05	Bias-Only Adaptation	Steering LLM Reasoning Through Bias-Only Adaptation
2025-11	Rank-1 LoRAs	Rank-1 LoRAs Encode Interpretable Reasoning Signals
2025-06	Layer Importance	Layer Importance for Mathematical Reasoning is Forged in Pre-Training and Invariant after Post-Training
2025-12	Effective Depth	What Affects the Effective Depth of Large Language Models?
2025-12	Internal Policy	Bottom-up Policy Optimization: Your Language Model Policy Secretly Contains Internal Policies
2025-06	The Emergence of Mutual Information (MI) Peaks	Demystifying Reasoning Dynamics with Mutual Information: Thinking Tokens are Information Peaks in LLM Reasoning

Unintended Behaviors

Hallucination

Click to hide/show the paper list

Date	Name	Title	Paper	Github
2025-05	FactEval	Are Reasoning Models More Prone to Hallucination?
2025-05	FSPO	Reasoning Models Hallucinate More: Factuality-Aware Reinforcement Learning for Large Reasoning Models
2025-05	RDH	Detection and Mitigation of Hallucination in Large Reasoning Models: A Mechanistic Perspective
2025-05	Hallucination Tax	The Hallucination Tax of Reinforcement Finetuning
2025-09	TTS_knowledge	Test-Time Scaling in Reasoning Models Is Not Effective for Knowledge-Intensive Tasks Yet
2025-05	Meta-Cognitive	Auditing Meta-Cognitive Hallucinations in Reasoning Large Language Models
2025-06	RACE	Joint Evaluation of Answer and Reasoning Consistency for Hallucination Detection in Large Reasoning Models

Unfaithfulness

Click to hide/show the paper list

Date	Name	Title	Paper	Github
2025-05	FactEval	Are Reasoning Models More Prone to Hallucination?
2025-01	Cue Insertion	Are DeepSeek R1 And Other Reasoning Models More Faithful?		-
2025-05	Don't Say What They Think	Reasoning Models Don't Always Say What They Think		-
2025-03	Unfaithful CoT	Chain-of-Thought Reasoning In The Wild Is Not Always Faithful
2025-06	Strategic Deception	When Thinking LLMs Lie: Unveiling the Strategic Deception in Representations of Reasoning Models
2025-05	Reasoning Intermediate Token	Beyond Semantics: The Unreasonable Effectiveness of Reasonless Intermediate Tokens		-
2025-07	Alignment Prediction	Can We Predict Alignment Before Models Finish Thinking? Towards Monitoring Misaligned Reasoning Models
2025-10	Refusal Cliff	Refusal Falls off a Cliff: How Safety Alignment Fails in Reasoning?		-
2025-03	CoT Monitor	Monitoring Reasoning Models for Misbehavior and the Risks of Promoting Obfuscation		-
2025-06	VFT	Teaching Models to Verbalize Reward Hacking in Chain-of-Thought Reasoning		-

Overthinking

Click to hide/show the paper list

Date	Name	Title	Paper	Github
2025-05	R1 Thoughtology	DeepSeek-R1 Thoughtology: Let’s think about LLM reasoning
2025-05	Between Underthinking and Overthinking	Between Underthinking and Overthinking: An Empirical Study of Reasoning Length and correctness in LLMs		-
2025-06	Does Thinking More Always Help?	Does Thinking More always Help? Mirage of Test-Time Scaling in Reasoning Models		-
2025-05	Short-m@k	Don't Overthink it. Preferring Shorter Thinking Chains for Improved LLM Reasoning		-
2025-02	Thinking-Optimal Scaling	Towards Thinking-Optimal Scaling of Test-Time Compute for LLM Reasoning
2024-12	Do NOT Think That Much	Do NOT Think That Much for 2+3=? On the Overthinking of o1-Like LLMs
2025-07	Inverse Scaling	Inverse Scaling in Test-Time Compute
2025-01	Underthinking	Thoughts Are All Over the Place: On the Underthinking of o1-Like LLMs		-
2025-02	ReasoningAction Dilemma	The Danger of Overthinking: Examining the Reasoning-Action Dilemma in Agentic Tasks		-
2025-10	REFRAIN	Stop When Enough: Adaptive Early-Stopping for Chain-of-Thought Reasoning		-
2025-05	AdaptThink	Stop When Enough: Adaptive Early-Stopping for Chain-of-Thought Reasoning
2025-05	Self-Braking Tuning	Let LRMs Break Free from Overthinking via Self-Braking Tuning
2025-04	MiP	Let LRMs Break Free from Overthinking via Self-Braking Tuning
2025-05	Behavioral Divergence	When Can Large Reasoning Models Save Thinking? Mechanistic Analysis of Behavioral Divergence in Reasoning		-
2025-04	Thought Manipulation	Thought Manipulation: External Thought Can Be Efficient for Large Reasoning Models		-
2025-07	Disregard Correct Solutions	Large Reasoning Models are not thinking straight: on the unreliability of thinking trajectories		-
2025-05	Manifold Steering	Mitigating Overthinking in Large Reasoning Models via Manifold Steering
2025-03	Representation	Towards Understanding Distilled Reasoning Models: A Representational Approach
2025-04	SEAL	SEAL: Steerable Reasoning Calibration of Large Language Models for Free
2025-05	Internal Bias	The First Impression Problem: Internal Bias Triggers Overthinking in Reasoning Models
2025-04	Self-Verification Probing	Reasoning Models Know When They're Right: Probing Hidden States for Self-Verification
2025-10	Zero-Step Thinking	The Zero-Step Thinking: An Empirical Study of Mode Selection as Harder Early Exit in Reasoning Models

Unsafety

Click to hide/show the paper list

Date	Name	Title	Paper	Github
2025-02	SafeChain	SafeChain: Safety of Language Models with Long Chain-of-Thought Reasoning Capabilities
2025-02	Hidden Risks	The Hidden Risks of Large Reasoning Models: A Safety Assessment of R1		-
2025-03	DeepSeek-R1-Safe	Safety Evaluation and Enhancement of DeepSeek Models in Chinese Contexts
2025-03	FreeEvalLM	Trade-offs in Large Reasoning Models: An Empirical Analysis of Deliberative and Adaptive Reasoning over Foundational Capabilities
2025-02	Mousetrap	A Mousetrap: Fooling Large Reasoning Models for Jailbreak with Chain of Iterative Chaos
2025-02	H-CoT	H-CoT: Hijacking the Chain-of-Thought Safety Reasoning Mechanism to Jailbreak Large Reasoning Models, Including OpenAI o1/o3, DeepSeek-R1, and Gemini 2.0 Flash Thinking
2025-08	R1-ACT	R1-ACT: Efficient Reasoning Model Safety Alignment by Activating Safety Knowledge
2025-10	COG	When Models Outthink Their Safety: Unveiling and Mitigating Self-Jailbreak in Large Reasoning Models

Training Methods

Combine SFT with RL

Click to hide/show the paper list

Date	Name	Title	Paper	Github
2025-09	Annealed-RLVR	How LLMs Learn to Reason: A Complex Network Perspective		-
2025-06	ReLIFT	Learning What Reinforcement Learning Can't: Interleaved Online Fine-Tuning for Hardest Questions		-
2025-12	TRAPO	Trust-Region Adaptive Policy Optimization
2025-05	UFT	UFT: Unifying Supervised and Reinforcement Fine-Tuning
2025-09	HPT	Towards a Unified View of Large Language Model Post-Training
2025-08	CHORD	On-Policy RL Meets Off-Policy Experts: Harmonizing Supervised Fine-Tuning and Reinforcement Learning via Dynamic Weighting
2025-06	SRFT	SRFT: A Single-Stage Method with Supervised and Reinforcement Fine-Tuning for Reasoning
2025-09	BRIDGE	Beyond Two-Stage Training: Cooperative SFT and RL for LLM Reasoning		-

RL balancing exploration and exploitation

Click to hide/show the paper list

Date	Name	Title	Paper	Github
2025-03	DAPO	DAPO: An Open-Source LLM Reinforcement Learning System at Scale
2025-05	Entropy Mechanism of RL	The Entropy Mechanism of Reinforcement Learning for Reasoning Language Models
2025-05	ProRL	ProRL: Prolonged Reinforcement Learning Expands Reasoning Boundaries in Large Language Models
2025-08	CURE	CURE: Critical-Token-Guided Re-Concatenation for Entropy-Collapse Prevention
2025-10	Reasoning Boundary Paradox	The Reasoning Boundary Paradox: How Reinforcement Learning Constrains Language Models
2025-09	outcome-based exploration	Outcome-based Exploration for LLM Reasoning		-
2025-06	entropy-based advantage	Reasoning with Exploration: An Entropy Perspective		-
2025-10	AEPO	Arbitrary Entropy Policy Optimization Breaks The Exploration Bottleneck of Reinforcement Learning
2025-09	AEnt	On Entropy Control in LLM-RL Algorithms		-
2025-09	SIREN	Rethinking Entropy Regularization in Large Reasoning Models