Reproduce “Toward Structured Knowledge Reasoning: Contrastive Retrieval-Augmented Generation on Experience” on OpenResearch

About this paper

Jiawei Gu, Ziting Xian, Yuanzhen Xie, Ye Liu, Enjie Liu, Ruichao Zhong, Mochi Gao, Yunzhi Tan and 2 more

The CoRE framework enhances Large Language Models' (LLMs) ability to reason with structured knowledge by building an 'experience memory' of diverse reasoning attempts and using contrastive retrieval-augmented generation. This approach yields average accuracy gains of 3.44% on Text-to-SQL and 3.4-4.25% on TableQA datasets.

Key insights

• Monte Carlo Tree Search effectively builds a dynamic 'Experience Memory' by generating diverse reasoning paths and intermediate steps, including partially successful ones, expanding the original training data 8-9 times.
• Integrating RAG with reward-informed and dynamic experience retrieval effectively guides LLMs in structured knowledge tasks by 'diluting intensive knowledge with effective examples'.
• Contrastive In-Context Learning (ICL), leveraging both positive and negative examples, consistently outperforms methods using only positive or fixed examples, demonstrating improved robustness and reasoning.

Summary by alphaXiv.