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

TL;DR

This work reframes Long Chain-of-Thought reasoning as a macromolecular structure governed by three bond-like interactions: Deep Reasoning (covalent), Self-Reflection (hydrogen bonds), and Self-Exploration (van der Waals). It demonstrates that robust Long CoT learning requires high-quality reasoning exemplars from strong reasoning LLMs, not surface keywords or human traces, and introduces Semantic Isomers to explain variability in learning outcomes. The authors propose Mole-Syn, a structure-aware data-synthesis framework, to transfer the bond topology from strong teachers to cheaper models, yielding improved Long CoT performance and RL stability across benchmarks. They also formalize attention-energy perspectives and bond-shaping functions, and show that synthetic bond distributions can be constructed from scratch while preserving desired reasoning dynamics. The approach highlights practical implications for data curation, model training, and defense against distillation of internal reasoning, with potential to enhance interpretability and reliability of long-horizon AI reasoning systems.

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.

Figures (16)

• Figure 1: The hypothesis that stable molecular structure in Long CoT arises from three key “chemical” bonds.
• Figure 2: Comparison of prior chain- or tree-like structures and our molecular structure. Reasoning starts from Mole. 0, uses deep reasoning on strongly related structures, then employs self-exploration for new logic in Mole. 1. When meet errors, reasoning utilize self-reflection to guide chain to optimized Mole. 0$^{+}$.
• Figure 3: The failure of distillation from weak instruction LLMs with ICL and Human-annotated reasoning traces to acquire Long CoT structures, compared to successful distillation from strong reasoning LLMs. See Appendix Figure \ref{['fig:comparison-full']} for the full result.
• Figure 4: Performance comparison between human-annotated reasoning traces (+ Human Distill Data) and R1 distilled reasoning traces (+ R1 Distill Data).
• Figure 5: Transfer graph on three different models. Pearson correlation coefficients across models are all greater than 0.9 (p<0.001), when sampling examples > 2,000, transfer graphs will become stable and get over 0.95 Pearson correlation between different sampling sizes.