Reproduce “Aligning Reasoning LLMs for Materials Discovery with Physics-aware Rejection Sampling” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Lee Hyun, Sohee Yoon, Jinwoo Park, Sue In Chae, Seongeon Park, Jooyeon Ahn, Yebin Jung, Youjung Chung and 6 more
Physics-aware Rejection Sampling (PaRS) is introduced to enhance the training of Large Reasoning Models (LRMs) for materials property prediction by generating and curating physically grounded reasoning traces. This approach leads to improved prediction accuracy and physical admissibility for recipe-to-property tasks in materials discovery, alongside increased computational efficiency in LRM supervision.
Key insights
- Incorporating domain-specific physical constraints directly into the rejection sampling process is crucial for curating high-quality reasoning traces that yield physically admissible and numerically accurate LRM predictions in scientific domains.
- Utilizing continuous error tolerance rather than binary correctness signals in LRM training significantly improves property prediction performance by better capturing the necessary numerical precision for scientific tasks.
- Adaptive halting mechanisms, based on trace diversity and error improvement, can substantially reduce the computational cost and time required for LRM supervision by efficiently pruning unproductive trace samples.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.