Presentation Information
[2Yin-A-40]A Reasoning Framework for Mathematical Optimization Support Integrating Step-Wise Logical Verification and Specification-Driven Implementation
〇Takumi Ito1, Tomonori Izumitani1 (1. NTT DOCOMO BUSINESS)
Keywords:
LLM,Multi-Agent,Mathematical Optimization
We present a multi-agent Large Language Model (LLM) approach for solving mathematical optimization problems. Despite the broad utility of optimization in decision-making, the requirement for high-level mathematical formulation and programming skills remains a major obstacle for non-specialists. Although current multi-agent systems attempt to automate these tasks, they often struggle with delayed validation and low specification compliance in code generation. Our proposed framework introduces two key innovations: an intermediate internal reflection agent that audits logic before implementation and a specification-driven code generation agent. Experimental results on benchmark datasets show that our approach outperforms existing baselines by increasing the success rate of optimal solution derivation and decreasing execution errors. These findings confirm that integrating early-stage logical reflection and strict specification adherence is vital for robust automated optimization.