LLMs Interpret Simulation Traces via Program Synthesis
A new arXiv paper (2602.10009v2) proposes an unsupervised learning method to translate fine-grained simulation traces into sparse, high-level pattern sequences for improved LLM interpretation. The approach uses program synthesis to create a library of pattern detectors, addressing scalability issues in tool-based LLM reasoning about physical systems. The work aims to enhance explainability and validation in LLMs' understanding of physics.
Key facts
- arXiv paper 2602.10009v2 proposes unsupervised learning for simulation trace annotation.
- Method translates simulation traces to sparse high-level pattern sequences.
- Uses program synthesis to create pattern detectors.
- Addresses scalability issues in LLM tooling for physical systems.
- Aims to improve LLM reasoning about specific physical systems.
- Current LLMs cannot reliably reason about physical systems.
- Explainability and validation remain open challenges.
- Tooling approach uses physical simulators for context.
Entities
Institutions
- arXiv