ThinkARM Framework Analyzes Reasoning in LLMs Using Schoenfeld's Theory
Researchers have unveiled a new scalable framework named ThinkARM (Anatomy of Reasoning in Models) designed to examine the reasoning patterns of large language models (LLMs). This framework, outlined in a paper on arXiv (ID: 2512.19995), utilizes Schoenfeld's Episode Theory as a lens to distill reasoning traces into functional components: Analyze, Explore, Implement, and Verify. When applied to mathematical problem-solving across various models, ThinkARM uncovers reproducible cognitive dynamics and highlights structural distinctions between reasoning and non-reasoning models that are not visible through token-level analysis. Two diagnostic case studies indicate that exploration serves as a vital branching point linked to correctness, while efficiency-focused methods tend to limit evaluative feedback steps. This research offers a fresh intermediate-scale perspective on LLM reasoning's cognitive framework.
Key facts
- ThinkARM framework abstracts reasoning traces into functional steps like Analysis, Explore, Implement, Verify.
- Based on Schoenfeld's Episode Theory.
- Applied to mathematical problem solving by diverse language models.
- Reveals structural differences between reasoning and non-reasoning models.
- Exploration is a critical branching step linked to correctness.
- Efficiency-oriented methods suppress evaluative feedback steps.
- Paper available on arXiv with ID 2512.19995.
- Framework is scalable and explicitly abstracts reasoning traces.
Entities
Institutions
- arXiv