Title: Is Neural-Integrated Mechanistic Modeling Ready for Large Language Models? Introducing a Benchmark and Agentic Framework

Abstract: While large language models (LLMs) have demonstrated potential in deriving mechanistic models from data, current evaluation methods predominantly rely on simplified scenarios, thereby overlooking the intricacies of actual scientific modeling. In real-world applications, this process frequently entails neural-integrated formulations, which require the joint construction of mechanistic and neural network components. This integration drastically expands the complexity of the search space. To bridge this gap, we present the Neural-Integrated Mechanistic Modeling (NIMM) benchmark, designed to assess LLM-generated models across three distinct scientific domains. Our experiments on NIMM indicate that conventional LLM-based methods face significant difficulties in navigating this complex landscape, leading to poor search stability and suboptimal solution quality. In response, we introduce NIMMGen, an agentic framework guided by tree structures. NIMMGen facilitates diverse exploration through branch-level search and enhances outcomes via atomic model refinement. Comprehensive testing confirms that NIMMGen delivers state-of-the-art results on the NIMM benchmark, markedly boosting both the stability of the search process and the quality of the final solutions.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC