SAGE: Sufficiency-Aware Implicit Graph Exploration for Long Context Reasoning

Zhangyu Wang, Siyuan Gao, Rong Zhou, Li Ning, Hao Wang

Keywords: Language Models, Knowledge Graph, Retrieval Augmented Generation

Abstract: Large Language Models (LLMs) have achieved impressive progress in natural language processing, but their limited ability to retain long-term context constrains performance on document-level or multi-turn tasks. Retrieval-Augmented Generation (RAG) mitigates this by retrieving relevant information from an external corpus. Recently, graph-based RAG systems have shown promise for long-context reasoning. However, these methods face some challenges: high preprocessing computational costs, static graph architectures with fixed node and edge semantics, and complex parameter finetuning requirements that could limit their practical adoption. Recognizing that modern LLMs possess substantially improved reasoning capabilities, we propose SAGE, a dynamic implicit graph exploration framework that eliminates the need for explicit graph construction while preserving multi-hop reasoning benefits. Experiments are conducted on challenging long-context QA benchmarks, including NovelQA and Marathon. Our approach consistently outperforms strong baselines across these datasets. Additionally, it reduces storage and runtime requirements by over an order of magnitude. These results show that high-quality retrieval can be achieved through LLM-driven text exploration without relying on static preprocessing or vector representations.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 23436