講演情報
[3B-03]Iterative LLM Augmentation through Criteria-Based Feedback
*宋 子瑜1、江口 浩二1 (1. 広島大学 データ解析・モデリング研究室)
発表者区分：学生
論文種別：ロングペーパー
インタラクティブ発表：あり
論文種別：ロングペーパー
インタラクティブ発表：あり
キーワード：
Large Language Model、Reflection、Retrieval-Augmented Generation、Prompt Engineering
Large Language Model (LLM) self-reflection involves an LLM reviewing its past outputs to enhance future responses without relying on external data. This concept has been explored in frameworks like those by Shinn et al. (2023)and Madaan et al. (2023). However, challenges remain, as Huang et al. (2024)point out the risk of performance degradation due to overly generic reflective prompts. To address these issues, we introduce a vector-based retrieval framework. Our approach demonstrates significant improvements in decision-making, reasoning, and mathematics tasks, surpassing baseline models like Llama3.2-3b and the SELF-REFINE framework. These results emphasize the potential of targeted self-reflection to improve LLM performance while mitigating common drawbacks. Meanwhile,beyond this method, we also explored the possibility of using a multi-agent approach with auxiliary models to assist in reflection. We trained a model to replace the base model in generating criteria and systematically evaluated the impact of the auxiliary model on the output capability of the self-reflection framework.