Self-GIVE: Associative Thinking from Limited Structured Knowledge for Enhanced Large Language Model Reasoning

2025-05-21

AI Summary

To address low efficiency, poor generalization, and high computational overhead in knowledge graph extrapolation—particularly when leveraging large language models (LLMs)—this paper proposes a reinforcement learning–driven retrieval-and-reasoning framework featuring a novel self-guided knowledge extrapolation mechanism. Unlike conventional approaches, it avoids explicit graph construction and triplet pruning, enabling efficient deployment of compact 3B/7B-parameter models. The method integrates retrieval-augmented generation (RAG), proximal policy optimization (PPO)-based reinforcement learning, structured information extraction, entity-set associative modeling, and lightweight UMLS integration. On unseen biomedical question-answering tasks, Qwen2.5-3B and Qwen2.5-7B achieve accuracies of 71.4% and 90.5%, respectively—surpassing baseline methods by over 40 percentage points. Moreover, token consumption is reduced by more than 90%, and the 7B variant outperforms GPT-3.5 Turbo augmented with GIVE, demonstrating strong scalability and inference efficiency.

Technology Category

• Data Mining & Knowledge Management: Conversational Systems for Recommendation & Retrieval
• Machine Learning: Large Multimodal Models (LMMs)
• Knowledge Representation and Reasoning: Computational Complexity of Reasoning

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Search and Retrieval-Augmented AI: Large language models for search

Abstract

When addressing complex questions that require new information, people often associate the question with existing knowledge to derive a sensible answer. For instance, when evaluating whether melatonin aids insomnia, one might associate"hormones helping mental disorders"with"melatonin being a hormone and insomnia a mental disorder"to complete the reasoning. Large Language Models (LLMs) also require such associative thinking, particularly in resolving scientific inquiries when retrieved knowledge is insufficient and does not directly answer the question. Graph Inspired Veracity Extrapolation (GIVE) addresses this by using a knowledge graph (KG) to extrapolate structured knowledge. However, it involves the construction and pruning of many hypothetical triplets, which limits efficiency and generalizability. We propose Self-GIVE, a retrieve-RL framework that enhances LLMs with automatic associative thinking through reinforcement learning. Self-GIVE extracts structured information and entity sets to assist the model in linking to the queried concepts. We address GIVE's key limitations: (1) extensive LLM calls and token overhead for knowledge extrapolation, (2) difficulty in deploying on smaller LLMs (3B or 7B) due to complex instructions, and (3) inaccurate knowledge from LLM pruning. Specifically, after fine-tuning using self-GIVE with a 135 node UMLS KG, it improves the performance of the Qwen2.5 3B and 7B models by up to $ extbf{28.5%$ ightarrow$71.4%}$ and $ extbf{78.6$ ightarrow$90.5%}$ in samples $ extbf{unseen}$ in challenging biomedical QA tasks. In particular, Self-GIVE allows the 7B model to match or outperform GPT3.5 turbo with GIVE, while cutting token usage by over 90%. Self-GIVE enhances the scalable integration of structured retrieval and reasoning with associative thinking.

Problem

• Enhancing LLM reasoning with associative thinking from limited knowledge
• Reducing LLM calls and token overhead in knowledge extrapolation
• Improving small LLM performance in complex biomedical QA tasks

Innovation

• Retrieve-RL framework for automatic associative thinking
• Extracts structured information to link queried concepts
• Reduces token usage and improves smaller LLMs performance