PaperlayerView exact revision on arXiv

Source-linked AI summary

LongSeeker: Elastic Context Orchestration for Long-Horizon Search Agents

Yijun Lu, Rui Ye, Yuwen Du, Jiajun Wang, Songhua Liu, Siheng Chen

arXiv:2605.05191v1cs.AI

TL;DR

Long-horizon search agents accumulate noisy, redundant context that can exceed what they retain. Context-ReAct jointly orchestrates reasoning, context edits, and tool use through five operations. LongSeeker outperforms named baselines on BrowseComp and BrowseComp-ZH.

  • Problem

    Long-horizon agents accumulate observations, reasoning traces, and tool calls, while existing approaches lack precise, on-demand control over the evolving context.

    Source: §1 · p. 2

  • Method

    Context-ReAct jointly generates reasoning, context operations, and tool calls at each turn, using five operations for multi-resolution context control.

    Source: §1 · pp. 2–3, §5 · p. 10

  • Results

    LongSeeker scores 61.5% on BrowseComp and 62.5% on BrowseComp-ZH, exceeding Tongyi DeepResearch (43.2%, 46.7%) and AgentFold (36.2%, 47.3%).

    Source: §1 · p. 2

  • Takeaways & Limitations

    Experiments validate more compact and efficient context management than append-only or coarse-grained truncation, while sustaining high performance at long horizons.

    Source: §5 · p. 10

    The current implementation uses supervised fine-tuning on synthesized trajectories without rejection sampling or advanced exploration strategies.

    Source: §5 · p. 10

Abstract

from arXiv · show

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.

1 Introduction

Long-horizon search agents face an increasingly noisy and oversized working context, while existing approaches lack precise, on-demand control. The paper proposes elastic context management through Context-ReAct and reports LongSeeker’s benchmark results.

1 Introduction

  • Existing long-horizon agents accumulate noisy, redundant context, while truncation, restarts, summarization, and curation approaches lack precise, on-demand control.The cited approaches are described as importance-agnostic, disruptive to continuity, fixed-granularity, error-accumulating, or limited in where and how they intervene.Source: §1 · p. 2
  • Elastic context management varies detail by task state, preserving fresh evidence, distilling resolved evidence, retaining critical snippets, and removing or rolling back failed branches.The goal is to maintain the appropriate fidelity for different parts of the agent’s reasoning trajectory.Source: §1 · p. 2
  • Context-ReAct jointly generates reasoning, context meta-operations, and the next tool call, applying context changes before appending the next observation.This lets the agent choose when and where to intervene and how prior context should be represented.Source: §1 · p. 2
  • Five operations—Skip, Compress, Snippet, Delete, and Rollback—preserve, summarize, quote, remove, or revert context at different fidelity levels.Compress is expressively complete; specialized operators provide efficiency and fidelity guarantees that reduce generation cost and hallucination risk.Source: §1 · p. 2
  • LongSeeker scores 61.5% on BrowseComp and 62.5% on BrowseComp-ZH, exceeding Tongyi DeepResearch (43.2%, 46.7%) and AgentFold (36.2%, 47.3%).LongSeeker is fine-tuned on 10k synthesized search trajectories.Source: §1 · p. 2

2 Related Work

Prior search agents use ReAct-style cycles but accumulate observations, while context-management methods offer different, incomplete forms of control. Context-ReAct combines content-aware operations with tool calls at each step.

2 Related Work

  • Conventional ReAct agents accumulate observations, which progressively degrades context quality and risks exceeding context windows during extended tasks.The ReAct paradigm structures behavior as iterative reasoning, action execution, and observation integration (Yao et al., 2023).Source: §2 · p. 3
  • Prior context-management methods use recency-based truncation, threshold-triggered flushing, fixed-buffer summarization, or proactive curation, which still lacks surgical edits and history revisiting.Examples include keep-last-k, discard-all, MEM1, MemAgent, AgentFold, and ARC.Source: §2 · p. 3
  • Context-ReAct co-generates five content-aware operations with tool calls at every step, spanning lossless extraction through structural backtracking.Its operations are invoked only when necessary and are learned end-to-end from synthesized long-horizon search trajectories.Source: §2 · p. 3

3 Method

Context-ReAct augments ReAct with an on-demand meta-action layer that lets agents reshape working context alongside reasoning and tool use. Its operations support context changes from exact extraction to structural backtracking.

3 Method

  • Context-ReAct adds an on-demand meta-action layer to ReAct, co-generating context edits alongside reasoning and tool calls.Source: §3 · p. 3, Figure 2 · p. 4

3.1 Agentic Paradigm

  • At each turn, the agent applies generated meta-operations before the next observation, keeping its working context compact and relevant without an external trigger.Source: §3.1 · p. 4

3.2 Atomic Meta-Operations

  • Compress summarizes any contiguous historical range, including nonrecent steps, while Skip leaves an already compact context unchanged.Source: §3.2 · p. 5
  • Rollback discards a failed later branch while preserving a summarized reason or transferable insight, and multiple operations can be composed in one step.Source: §3.2 · p. 5
  • Snippet extracts an exact anchored substring, whereas Delete removes an entire uninformative step to reduce noise.Source: §3.2 · p. 5

3.3 Expressive Completeness and Principled Redundancy

  • Compress alone is expressively complete: one operation can transform any input history into any target history.Source: §3.3 · p. 6
  • The specialized operators provide practical structure for efficient, reliable context management, with Snippet preserving exact content that generative compression may distort.Source: §3.3 · pp. 6–7

3.4 Data Synthesis and Training

  • The training pipeline builds 10,000 annotated trajectories from multi-hop questions and teacher rollouts, then fine-tunes Qwen 3 30B-A3B on their structured outputs.Source: §3.4 · p. 7

4 Experiments

LongSeeker is evaluated on four search benchmarks and against diverse agent baselines. Its results accompany evidence that Context-ReAct keeps context compact, uses all five meta-operations, and outperforms alternative context-management strategies under the same step budget.

4.1 Experimental Setup

  • LongSeeker is evaluated on BrowseComp, BrowseComp-ZH, xbench, and GAIA-text, with up to 300 tool calls; BrowseComp and BrowseComp-ZH use 200 sampled questions each.The baseline set includes foundation models with tools and comparable-scale search agents.Source: §4.1 · pp. 7–8

4.2 Results and Analysis

  • 61.5 on BrowseComp and 62.5 on BrowseComp-ZH place LongSeeker ahead of Tongyi-DeepResearch (43.4) and AgentFold (36.2) on BrowseComp, and it scores 78.0 on xbench-2505 and 77.7 on GAIA-text.Table 1’s caption also reports that LongSeeker outperforms GPT-5 and Gemini-3.0-Pro on BrowseComp despite having 30B parameters.Source: §4.2 · p. 8, Table 1 · p. 8
  • LongSeeker’s context stays under 15k tokens even at 300-step horizons, while ReAct-based DeepSeek-V3.2 exhibits rapid, unbounded context growth.The managed context uses only a fraction of the 256k context window; the paper attributes this compactness to its meta-operations.Source: §4.2 · p. 9, Figure 4 · p. 9
  • LongSeeker uses all five meta-operations across trajectories, with Snippet and Delete used less often as early-stage uncertainty encourages conservative pruning.The usage distribution is shown in Figure 4b.Source: §4.2 · p. 9
  • Under the same step budget and base-model configuration, Context-ReAct achieves the highest BrowseComp performance among Context-ReAct, Summary, and Discard-all.Summary compresses overflowed trajectories; Discard-all removes previous tool-call history.Source: §4.2 · pp. 9–10, Figure 5 · p. 10

5 Conclusion

Context-ReAct coordinates reasoning, context edits, and tool calls through five atomic operations for fine-grained control of working context. LongSeeker demonstrates competitive benchmark performance, while the current implementation leaves reinforcement learning and advanced exploration for future work.

5 Conclusion

  • Context-ReAct jointly generates reasoning, context meta-operations, and tool calls at each step, using five atomic operations for fine-grained control of evolving context.The operations are Skip, Compress, Rollback, Snippet, and Delete.Source: §5 · p. 10
  • LongSeeker demonstrates competitive long-horizon search benchmark performance and more compact context management than append-only or coarse-grained truncation approaches.The conclusion notes that it surpasses Tongyi DeepResearch and AgentFold on BrowseComp.Source: §5 · p. 10
  • The current implementation uses SFT on synthesized trajectories without rejection sampling or advanced exploration; future work includes applying RL to optimize meta-operation usage.The authors envision Context-ReAct as a domain-agnostic paradigm for other long-horizon challenges.Source: §5 · p. 10

A Case Study

The case study shows managed context after meta-operations and the corresponding structured output from LongSeeker.

A Case Study

Related papers

Loading 2605.05191v1…