Presentation Information
[4Yin-B-47]Principles-driven Task-Specific Large Reasoning ModelsPilot Study on System Design Document Review
〇Yuta Koreeda1, Koichi Nagatsuka1, Kazu Nishikawa1, Yasuhide Mori1, Shun Oida1, Daisuke Fukui1 (1. Hitachi, Ltd.)
Keywords:
System engineering,Natural language processing,Large language model
This study proposes a data-efficient method for developing task-specific large reasoning models (LRM) in micro-domains. The proposed method relies on extracting fundamental reasoning principles from a small number of examples and using them to guide supervised fine-tuning. We evaluate the proposed method on checklist-based reviews of system design documents, formulating the task as generation of pass/fail judgments and corresponding rationales for each checklist item. Experimental results show that incorporating extracted principles improves the accuracy of a Qwen3-8B model from 0.750 to 0.767 without increasing inference cost. While some commercial models achieve higher accuracy, they incur substantially higher cost, indicating that the proposed approach offers a practical trade-off between accuracy and efficiency for enterprise applications.