Work

ARC: Active and Reflection-driven Context Management for Long-Horizon Information Seeking Agents

Year: 2026

Type: conference-paper

Abstract: Large language models are increasingly deployed as research agents for deep search and long-horizon information seeking, yet their performance often degrades as interaction histories grow.This degradation, known as context rot, reflects a failure to maintain coherent and task-relevant internal states over extended reasoning horizons.Existing approaches primarily manage context through raw accumulation or passive summarization, treating it as a static artifact and allowing early errors or misplaced emphasis to persist.Motivated by this perspective, we propose ARC, an active and reflectiondriven framework that formulates context management as the online maintenance of a dynamically revisable internal reasoning state during execution.ARC operationalizes this view through incremental context construction together with selectively triggered reflection, allowing agents to reorganize their working context when misalignment or degradation is detected.Experiments on five challenging longhorizon information-seeking benchmarks show that ARC consistently outperforms raw history accumulation and passive summarization baselines, achieving up to an 11% absolute improvement on BrowseComp-ZH with Qwen2.5-32B-Instruct. 1 rather than revised in light of downstream reasoning outcomes.More fundamentally, the challenge of long-horizon information seeking lies not only in deciding what to retain, but also in ensuring that the maintained context can evolve as the agent's understanding changes.

Source: Findings of the Association for Computational Linguistics: ACL (Findings of ACL)

DOI: 10.18653/v1/2026.findings-acl.930