The Molecular Structure of Thought: Mapping the Topology of Long Chain-of-Thought Reasoning

Qiguang Chen, Yantao Du, Ziniu Li, Jinhao Liu, Songyao Duan, Jiarui Guo, Minghao Liu, Jiaheng Liu, Tong Yang, Ge Zhang, Libo Qin, Wanxiang Che, Wenhao HuangarXiv 2026 · arXiv · OpenReview

TL;DR

This paper proposes a novel "molecular structure" analogy to understand how Large Language Models (LLMs) learn complex, multi-step reasoning (Long Chain-of-Thought or Long CoT). They introduce a method, MOLE-SYN, to synthesize these reasoning structures, significantly improving LLM performance and stability, especially in reinforcement learning.

What Problem Does This Paper Address?

LLMs are great at many tasks, but they often struggle with "Long Chain-of-Thought" (Long CoT) reasoning, which involves a long, coherent sequence of intermediate steps to reach a complex conclusion. While humans can generate these long reasoning chains naturally, LLMs often fail to learn them effectively, even when trained on human examples or distilled from other models. This gap limits their ability to tackle intricate problems requiring deep, multi-step logical deduction. The authors aim to understand why LLMs fail to learn Long CoT and propose a new way to teach them.

Methods

The core of the paper is the "molecular structure" analogy for Long CoT. They propose that effective Long CoT reasoning isn't just a linear sequence of steps but a structured "molecule" formed by three types of "bonds" representing different reasoning behaviors:

  • Deep Reasoning (Covalent Bonds): Forms the backbone, representing strong logical dependencies and deductive steps.
  • Self-Reflection (Hydrogen Bonds): Acts as a stabilizer, allowing the reasoning to fold back on itself, test assumptions, and correct errors.
  • Self-Exploration (Van der Waals Forces): Bridges distant concepts, allowing for hypothesis generation and exploring alternative paths.

They developed MOLE-SYN, a method that uses these "bond distributions" to synthesize effective Long CoT structures. Instead of just copying reasoning steps, MOLE-SYN learns the underlying behavioral transition graph from strong reasoning models and uses it to guide instruction-tuned LLMs to generate coherent Long CoT trajectories from scratch. This approach decouples the reasoning structure from the specific wording, making it more robust.

Datasets

The paper uses several datasets for training and evaluation:

  • OpenThoughts-3: A large corpus of Long CoT queries used for generating synthetic training data.
  • Mathematical Reasoning Benchmarks:
    • GSM8K (grade-school math)
    • MATH-500 (high-school math)
    • AMC 2023 (high-school competitions)
    • AIME 2024/2025 (invitational math)
    • OlymBench (comprehensive Olympiad-level benchmark)
  • Distilled Data: Data distilled from strong reasoning LLMs (e.g., DeepSeek-R1, OpenAI-OSS, QwQ-32B) and weak instruction LLMs.

Key Insights and Ideas

  • Long CoT is a "Molecular Structure": The most significant insight is that effective Long CoT reasoning has a stable, molecular-like topology, not just a linear sequence. The interplay of Deep Reasoning, Self-Reflection, and Self-Exploration bonds is crucial for its stability and learnability.
  • Structure Over Keywords: LLMs learn the underlying reasoning structure and behaviors, not just the specific keywords used in reasoning traces. This explains why simply imitating keywords from weaker models doesn't lead to effective Long CoT.
  • Semantic Isomers: Different Long CoT trajectories can solve the same problem but have different "bond structures." These "semantic isomers" highlight that the arrangement of reasoning behaviors is key, not just the concepts involved. Mixing incompatible isomers leads to performance degradation.
  • MOLE-SYN: Synthesizing Reasoning Structures: The MOLE-SYN method demonstrates that it's possible to synthesize effective Long CoT structures from scratch using instruction-tuned LLMs by learning the behavioral transition graph, offering a more efficient way to transfer reasoning capabilities.
  • Metacognitive Oscillation: LLMs exhibit a "metacognitive oscillation" between exploration and validation, which is different from human reasoning. Aligning training objectives with these behavioral structures can bridge this gap.

Key Results

  • MOLE-SYN Boosts Performance: Models trained with MOLE-SYN generated data achieve performance close to or exceeding distillation from strong reasoning LLMs, even when using weaker instruction-tuned models. For example, on the QwQ benchmark, MOLE-SYN achieved performance comparable to QwQ distillation.
  • Improved RL Stability: MOLE-SYN-initialized models show significantly better and more sustained performance gains in reinforcement learning (RL) tasks compared to models initialized from base LLMs or standard distillation.
  • Structural Compatibility is Crucial: Simply having similar statistical distributions of reasoning traces is not enough. Co-activating incompatible reasoning structures leads to "structural chaos" and performance degradation, underscoring the importance of structural alignment.
  • Bond Distributions Matter: The paper quantifies the energy levels associated with different bond types, showing that lower-energy bonds (like Deep Reasoning) are favored and act as constraints, stabilizing the reasoning process.

Limitations and Future Work

  • Limited Teacher Models and Backbones: The analysis relied on a specific set of teacher models and LLM backbones, which might introduce biases.
  • Offline Distillation Focus: The current work focuses on offline distillation and supervised fine-tuning. Future work could explore how these methods scale in online or interactive settings with RL feedback.
  • Approximate Visualization: Visualizing the geometric characteristics of inferred bonds in information and semantic spaces is approximate. Developing more precise methods for delineating universal Long CoT structures is a future direction.
  • Automatic Labeling: The behavior analysis relies on automatically labeled datasets, which might contain noise or bias. Initial robustness checks were performed, but further validation is needed.
  • Human vs. LLM Reasoning Dynamics: While the paper highlights differences in information flow and metacognitive oscillation between humans and LLMs, further work is needed to fully bridge this gap.

Summary updated

Abstract

Large language models (LLMs) often fail to learn effective long chain-of-thought (Long CoT) reasoning from human or non-Long-CoT LLMs imitation. To understand this, we propose that effective and learnable Long CoT trajectories feature stable molecular-like structures in unified view, which are formed by three interaction types: Deep-Reasoning (covalent-like), Self-Reflection (hydrogen-bond-like), and Self-Exploration (van der Waals-like). Analysis of distilled trajectories reveals these structures emerge from Long CoT fine-tuning, not keyword imitation. We introduce Effective Semantic Isomers and show that only bonds promoting fast entropy convergence support stable Long CoT learning, while structural competition impairs training. Drawing on these findings, we present Mole-Syn, a distribution-transfer-graph method that guides synthesis of effective Long CoT structures, boosting performance and RL stability across benchmarks.
BibTeX
@misc{chen2026molecular,
  author = {Qiguang Chen and Yantao Du and Ziniu Li and Jinhao Liu and Songyao Duan and Jiarui Guo and Minghao Liu and Jiaheng Liu and Tong Yang and Ge Zhang and Libo Qin and Wanxiang Che and Wenhao Huang},
  title = {{The Molecular Structure of Thought: Mapping the Topology of Long Chain-of-Thought Reasoning}},
  year = {2026},
  eprint = {2601.06002},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2601.06002},
}
Figure