Self-Taught Agentic Long Context Understanding

TL;DR

AgenticLU enhances large language models' ability to understand complex, long-context questions by integrating self-clarification and contextual grounding, leading to improved reasoning and retrieval efficiency across multiple tasks.

Contribution

The paper introduces AgenticLU, a novel framework that combines self-clarification with contextual grounding and a tree search inference method, improving long-context understanding in LLMs.

Findings

Achieves 97.8% answer recall on NarrativeQA with limited search depth.

Outperforms state-of-the-art prompting methods and specialized long-context models.

Maintains robust multi-hop reasoning as context length increases.

Abstract

Answering complex, long-context questions remains a major challenge for large language models (LLMs) as it requires effective question clarifications and context retrieval. We propose Agentic Long-Context Understanding (AgenticLU), a framework designed to enhance an LLM's understanding of such queries by integrating targeted self-clarification with contextual grounding within an agentic workflow. At the core of AgenticLU is Chain-of-Clarifications (CoC), where models refine their understanding through self-generated clarification questions and corresponding contextual groundings. By scaling inference as a tree search where each node represents a CoC step, we achieve 97.8% answer recall on NarrativeQA with a search depth of up to three and a branching factor of eight. To amortize the high cost of this search process to training, we leverage the preference pairs for each step obtained by the CoC workflow and perform two-stage model finetuning: (1) supervised finetuning to learn effective decomposition strategies, and (2) direct preference optimization to enhance reasoning quality. This enables AgenticLU models to generate clarifications and retrieve relevant context effectively and efficiently in a single inference pass. Extensive experiments across seven long-context tasks demonstrate that AgenticLU significantly outperforms state-of-the-art prompting methods and specialized long-context LLMs, achieving robust multi-hop reasoning while sustaining consistent performance as context length grows.