LongSeeker: Elastic Context Orchestration for Long-Horizon Search Agents

Problem
This paper addresses the challenge of managing context in long-horizon search agents, particularly the inefficiencies and error rates associated with naively accumulating all intermediate content during reasoning and tool use. The authors argue that existing literature lacks effective strategies for adaptive context management, which is crucial for maintaining performance as the working context grows. This work is presented as a preprint and has not yet undergone peer review.

Method
The core technical contribution is the introduction of Context-ReAct, a novel paradigm for elastic context orchestration that integrates reasoning, context management, and tool utilization in a unified framework. Context-ReAct comprises five atomic operations: Skip, Compress, Rollback, Snippet, and Delete. These operations allow agents to dynamically adjust their working context based on task relevance, enabling them to preserve critical information, summarize resolved data, discard irrelevant branches, and manage context size effectively. The authors demonstrate that the Compress operator is expressively complete, while the other operators provide guarantees for efficiency and fidelity, thereby reducing generation costs and minimizing hallucination risks. LongSeeker, the specific agent developed under this paradigm, is fine-tuned from the Qwen3-30B-A3B model using 10,000 synthesized trajectories.

Results
LongSeeker was evaluated across four representative search benchmarks, achieving notable performance metrics: 61.5% on BrowseComp and 62.5% on BrowseComp-ZH. These results significantly surpass those of the baseline models, Tongyi DeepResearch (43.2% and 46.7%) and AgentFold (36.2% and 47.3%). The effect sizes indicate a substantial improvement in long-horizon reasoning capabilities, underscoring the efficacy of adaptive context management in enhancing agent performance.

Limitations
The authors acknowledge several limitations, including the reliance on synthesized trajectories for training, which may not fully capture the complexities of real-world scenarios. Additionally, while the Context-ReAct framework is designed to be general, its performance in highly dynamic or unpredictable environments remains untested. The paper does not address potential scalability issues when applied to larger or more complex tasks, nor does it explore the implications of context management on the interpretability of agent decisions.

Why it matters
The implications of this work are significant for the development of more efficient and reliable long-horizon search agents. By demonstrating that adaptive context management can enhance reasoning capabilities, this research paves the way for future investigations into context orchestration techniques. It opens avenues for improving agent performance in various applications, including natural language processing, robotics, and decision-making systems, where managing large amounts of contextual information is critical.

Authors: Yijun Lu, Rui Ye, Yuwen Du, Jiajun Wang, Songhua Liu, Siheng Chen
Source: arXiv:2605.05191
URL: https://arxiv.org/abs/2605.05191v1