Aligning Reasoning LLMs for Materials Discovery with Physics-aware Rejection Sampling
TLDR
Introduces Physics-aware Rejection Sampling (PaRS) to train reasoning LLMs for accurate, calibrated, and physically admissible property prediction in materials discovery.
Reasoning
Strengths: Novel domain-aware training method (PaRS) that improves accuracy, calibration, and reduces physics violations; clear methodology and evaluation against baselines. Weaknesses: Limited to property prediction tasks; no real-world experimental validation or deployment in closed-loop systems is demonstrated.
Read-first score
Read-first score 46.1, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 40.
Field roles
Rank sensitivity
Stability: volatile; rank range: 10.
Keyword Scores
Deep Analysis
Innovations
- Physics-aware Rejection Sampling (PaRS) for training-time selection of reasoning traces
- Incorporation of physical admissibility (consistency with fundamental physics and numerical closeness) into trace selection criteria
- Lightweight halting mechanism to control compute during trace selection
Methodology
A teacher-student framework where a larger teacher model generates reasoning traces, and PaRS selects traces that are physically consistent and numerically close to targets, with halting to limit compute. The student model is fine-tuned on these selected traces and evaluated under matched token budgets against rejection sampling baselines.
Key Results
PaRS improves accuracy and calibration, reduces physics-violation rates, and lowers sampling cost compared to baselines using binary correctness or learned preference signals.