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

TL;DR

Long-horizon information-seeking with LLMs suffers from context rot as interaction histories grow. The authors propose ARC, an active, reflection-driven context-management framework that treats context as a dynamic internal state maintained by a dedicated Context Manager decoupled from action generation. ARC uses always-on incremental interaction memory and reflection-triggered revision to repair degraded context, with a training regime based on supervised fine-tuning of the CM. Across five benchmarks, ARC outperforms passive context compression baselines, with up to 11% absolute accuracy gains, and a trained CM can surpass larger untrained models while reducing memory cost. This work suggests that principled context management is a learnable, crucial component for robust long-horizon reasoning in information-seeking agents.

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, which is the first framework to systematically formulate context management as an active, reflection-driven process that treats context as a dynamic internal reasoning state during execution. ARC operationalizes this view through reflection-driven monitoring and revision, allowing agents to actively reorganize their working context when misalignment or degradation is detected. Experiments on challenging long-horizon information-seeking benchmarks show that ARC consistently outperforms passive context compression methods, achieving up to an 11% absolute improvement in accuracy on BrowseComp-ZH with Qwen2.5-32B-Instruct.

Figures (5)

• Figure 1: High-level illustration of active and reflection-driven context management in ARC. At each turn, the Context Manager actively updates the working context, enabling the actor to adapt its reasoning as understanding evolves across turns.
• Figure 2: Overview of the Context Manager in ARC. The upper panel illustrates standard summarization, which compresses accumulated interactions into a compact interaction memory. The lower panel shows reflection-driven context revision, where the Context Manager actively reorganizes the checklist and interaction memory to correct errors or misalignment, producing an updated execution-time context.
• Figure 3: The execution loop of ARC. At each turn, the actor generates reasoning and actions conditioned on the current context, producing an interaction consisting of thoughts, actions, and observations. The Context Manager (CM) then integrates the query, checklist, interaction memory, and the most recent interaction to update the context via active context management, which is passed to the next turn until a final answer is produced.
• Figure 4: Interaction efficiency under varying maximum interaction turns. Accuracy denotes the average performance on xBench-DeepSearch and BrowseComp-ZH, reflecting how effectively each method utilizes additional interaction budget.
• Figure 5: Token efficiency on BrowseComp. Context token usage is averaged over trajectories as interaction length increases. ARC maintains a compact interaction memory, while raw interaction history grows nearly linearly.