LongSeeker: Elastic Context Orchestration for Long-Horizon Search Agents
Paper: https://arxiv.org/abs/2605.05191
Authors: Yijun Lu, Rui Ye, Yuwen Du, Jiajun Wang, Songhua Liu, Siheng Chen
Overview
Long-horizon search agents must manage a rapidly growing working context as they reason, call tools, and observe information. Naively accumulating all intermediate content can overwhelm the agent, increasing costs and the risk of errors. The paper proposes that effective context management should be adaptive: parts of the agent's trajectory are maintained at different levels of detail depending on their current relevance to the task.
Context-ReAct Paradigm
To operationalize this principle, the authors introduce Context-ReAct, a general agentic paradigm for elastic context orchestration that integrates reasoning, context management, and tool use in a unified loop. It provides five atomic operations:
- Skip — bypass unhelpful content
- Compress — summarize resolved information
- Rollback — revert to earlier states
- Snippet — preserve key evidence
- Delete — discard unhelpful branches and control context size
- Adaptive context management—actively shaping working memory rather than passively accumulating it—yields more reliable and efficient long-horizon reasoning.
- Formal guarantees (expressive completeness of Compress, efficiency/fidelity of specialized operators) connect context manipulation to reduced cost and hallucination risk.
- Synthesized trajectories (10k) suffice to fine-tune a strong agent on top of Qwen3-30B-A3B.
- A Systematic Framework for Enterprise Knowledge Retrieval
- Retrieval Augmented Generation and Understanding in Vision: A Survey
- Synergizing RAG and Reasoning: A Systematic Review
- AceSearcher: Bootstrapping Reasoning and Search for LLMs via Reinforcement Learning
- Agentic Information Retrieval
These operations let the agent dynamically reshape its working context. The authors prove that the Compress operator is expressively complete, while the other specialized operators provide efficiency and fidelity guarantees that reduce generation cost and hallucination risk.
LongSeeker Model and Results
Building on this paradigm, LongSeeker is a long-horizon search agent fine-tuned from Qwen3-30B-A3B on 10k synthesized trajectories. Across four representative search benchmarks:
| Benchmark | LongSeeker | Tongyi DeepResearch | AgentFold | |-----------|-----------|---------------------|-----------| | BrowseComp | 61.5% | 43.2% | 36.2% | | BrowseComp-ZH | 62.5% | 46.7% | 47.3% |
Key Takeaways
Limitations and Discussion
Potential limitations include benchmark-vs-real-user distribution gaps, security risks of agents operating on the open web, and compute constraints on experimentation scale. Future directions may include more efficient test-time compute allocation, deeper integration with knowledge graphs, and cross-lingual/multimodal extension.
Original Abstract
> Long-horizon search agents must manage a rapidly growing working context as they reason, call tools, and observe information. Naively accumulating all intermediate content can overwhelm the agent, increasing costs and the risk of errors. We propose that effective context management should be adaptive: parts of the agent's trajectory are maintained at different levels of detail depending on their current relevance to the task. To operationalize this principle, we introduce Context-ReAct, a general agentic paradigm for elastic context orchestration that integrates reasoning, context management, and tool use in a unified loop. Context-ReAct provides five atomic operations: Skip, Compress, Rollback, Snippet and Delete, which allow the agent to dynamically reshape its working context, preserving important evidence, summarizing resolved information, discarding unhelpful branches, and controlling context size. We prove that the Compress operator is expressively complete, while the other specialized operators provide efficiency and fidelity guarantees that reduce generation cost and hallucination risk. Building on this paradigm, we develop LongSeeker, a long-horizon search agent fine-tuned from Qwen3-30B-A3B on 10k synthesized trajectories. Across four representative search benchmarks, LongSeeker achieves 61.5% on BrowseComp and 62.5% on BrowseComp-ZH, substantially outperforming Tongyi DeepResearch (43.2% and 46.7%) and AgentFold (36.2% and 47.3%). These results highlight the potential of adaptive context management, showing that agents can achieve more reliable and efficient long-horizon reasoning by actively shaping their working memory.