LLM Reasoning Enhanced with Scientific Logicality Methodology in Physics

A new methodology introduces logicality-enriched training for Large Language Models (LLMs) to improve scientific reasoning, specifically in physics. Current approaches focus on boosting performance on QA benchmarks through larger datasets and extended reasoning chains, but neglect the logical foundation essential for valid reasoning steps. This work systematically investigates internal logicality in LLM scientific reasoning, developing assessment criteria and data sampling methods for logicality-guided training. The goal is to enhance logical faithfulness and task performance. The research is published on arXiv with ID 2605.17104.