LongSeeker: Elastic Context Orchestration for Long-Horizon Search Agents Yijun Lu1,* , Rui Ye1,*,#,† , Yuwen Du1 , Jiajun Wang1 , Songhua Liu1,† , Siheng Chen1,† 1 Shanghai Jiao Tong University, * Equal Core Contributions, # Project Lead † Corresponding Authors: {yr991129, liusonghua, sihengc}@sjtu.edu.cn
arXiv:2605.05191v1 [cs.AI] 6 May 2026
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. Code Model
https://github.com/PolarSeeker/LongSeeker https://huggingface.co/PolarSeeker/LongSeeker-30B-SFT
Figure 1: LongSeeker-30B delivers strong results on challenging long-horizon benchmarks, matching or surpassing several foundation models and search agents.
Preprint.
1
Introduction
The emergence of search agents has transformed how humans retrieve and synthesize information from the web. Tasks once requiring manual query iteration can now be delegated end-to-end to an AI agent via a single instruction. Agentic search has thus become a cornerstone capability pursued by AI labs, exemplified by the trajectory from OpenAI’s Deep Research systems (OpenAI, 2025) to today’s top-tier large language models supporting multi-step, tool-augmented reasoning built upon the ReAct paradigm (Yao et al., 2023). However, long-horizon search agents built on the ReAct paradigm face an inherent context bottleneck: as observations, reasoning traces, and tool calls accumulate, the working context becomes increasingly noisy, redundant, and eventually too long to retain in full. Existing remedies remain partial: sliding-window truncation is importance-agnostic (Team et al., 2026); threshold-triggered re-starting disrupts reasoning continuity (DeepSeek-AI, 2025); periodic summarization suffers from fixed-granularity compression and accumulating abstraction errors (Zhou et al., 2025c; Yu et al., 2025; Lu et al., 2025); and proactive curation is still limited in where or how it can intervene (Ye et al., 2025; Yao et al., 2026). As a result, existing methods cannot provide precise, on-demand control over the evolving shape of the agent’s context. Addressing these, our key insight is that effective context management requires an elastic working context, one that can compress, preserve, discard, and restructure different parts of the agent’s context according to its current state. That is, during long-horizon search, information should exist in different forms as the task evolves: fresh evidence may need to remain intact for verification, resolved evidence can be distilled into conclusions, precision-critical details may survive as snippets, and failed branches should be removed or rolled back. This state-dependent fidelity ensures that the agent maintains the right level of detail for each part of its agentic reasoning trajectory. Following this insight, we propose Context-ReAct, a general agentic paradigm for elastic context orchestration in long-horizon search agents. At each reasoning turn, the agent jointly produces its reasoning trace, a set of context meta-operations, and the next tool call in a single autoregressive pass. The meta-operations are applied before the next observation is appended, allowing the agent to actively determine when to update its context, where in the trajectory to intervene, and how each part of the past should be represented. Based on this paradigm, we design a meta-operation vocabulary consisting of five atomic actions. (1) Skip leaves the context unchanged when it is already compact and informative. (2) Compress replaces any contiguous range of historical steps with an abstractive summary. (3) Snippet preserves an exact substring from a step, retaining precision-critical evidence such as numbers, entity names, quotations, or code without abstractive distortion. (4) Delete removes a step which no longer carries residual value. (5) Rollback abandons an unproductive branch by reverting the context to an earlier state while recording the reason for backtracking and any transferable insight. Together, these operations maintain a multi-resolution working context in which different parts of the context can remain verbatim, be compressed, be partially quoted, be removed, or be structurally rolled back. Although simple, this operation set is proved to be expressively complete: C OMPRESS alone can simulate every other operation in principle, while the specialized operators reduce generation cost and hallucination risk through explicit efficiency and fidelity guarantees. To instantiate and evaluate Context-ReAct, we build LongSeeker, a long-horizon search agent finetuned from Qwen3-30B-A3B on 10k synthesized search trajectories. We evaluate LongSeeker on four representative search benchmarks: BrowseComp (Wei et al., 2025b), BrowseComp-ZH (Zhou et al., 2025a), xbench (Chen et al., 2025), and GAIA (Mialon et al., 2023). LongSeeker achieves scores of 61.5% and 62.5% on BrowseComp and BrowseComp-ZH respectively, significantly outperforming competitive baselines such as Tongyi DeepResearch (43.2% and 46.7%) (Team et al., 2025) and AgentFold (36.2% and 47.3%) (Ye et al., 2025). These results suggest that elastic context orchestration is a scalable path toward more capable long-horizon agents, shifting context management from a peripheral engineering heuristic to a core component of agentic reasoning.
2
Our main contributions are: • Paradigm. We propose Context-ReAct, a general agentic paradigm for elastic context orchestration that lets agents decide when, where, and how to reshape their working context during ReAct-style search. • Operations. We introduce five meta-operations, Skip, Compress, Rollback, Snippet, and Delete, forming an expressively complete yet efficient operation set for multi-resolution context control. • Experiments. We train LongSeeker on 10k synthesized trajectories and achieve 61.5% on BrowseComp and 62.5% on BrowseComp-ZH, outperforming strong long-horizon search baselines.
2
Related Work
Search Agents. LLM-based search agents have transformed information retrieval from static queryresponse matching into dynamic, multi-step reasoning processes. Central to this transformation is the ReAct paradigm (Yao et al., 2023), which structures agent behavior as an iterative cycle of reasoning, action execution, and observation integration. OpenAI’s Deep Research (OpenAI, 2025a) pioneers the fully closed-source path, followed by a series of proprietary agents; meanwhile, open-source efforts such as WebSailor (Li et al., 2025) and Tongyi DeepResearch (Team et al., 2025) push capabilities forward through large-scale trajectory synthesis and post-training optimization. Yet these advances retain a fundamental limitation: they follow the conventional ReAct pattern of unconditionally accumulating observations, causing progressive degradation in context quality and heightened risk of exceeding context windows during extended tasks. Context Management for Agents. Managing growing context has attracted considerable recent attention, with existing methods falling into four categories. Sliding-window heuristics such as keep-last-k —adopted by the MiroThinker series (Team et al., 2026)—retain only recent steps and discard older content regardless of importance, while discard-all variants flush the entire context once thresholds are reached, as in DeepSeek-V3.2 (DeepSeek-AI, 2025) and GLM-4.7 (Zhipu AI, 2025). Periodic summarization methods such as MEM1 (Zhou et al., 2025c) train agents to maintain compact internal states and achieve strong multihop QA performance, while MemAgent (Yu et al., 2025) processes documents in segments with fixed-size memory buffers. Proactive curation approaches such as AgentFold (Ye et al., 2025) and ARC (Yao et al., 2026) enable agents to actively decide what and when to compress, yet they still lack surgical operations and cannot revisit earlier history to purge outdated content. Our Approach. Context-ReAct advances beyond prior work by defining a formally complete set of five atomic meta-operations—Skip, Compress, Rollback, Snippet and Delete—that are co-generated with standard tool calls at every step. Unlike fixed-rule truncation, each decision is content-aware; unlike periodic summarization, operations are invoked only when necessary; unlike proactive curation, our operation set is formally proven complete (Section 3) and spans the full spectrum from lossless extraction to structural backtracking. All decisions are learned end-to-end from synthesized long-horizon search trajectories.
3
Method
We present Context-ReAct, a general paradigm that augments the standard ReAct loop with an explicit meta-action layer for on-demand context management. As illustrated in Figure 2, our approach enables the agent to actively curate its working memory by co-generating meta-operations alongside standard reasoning and tool calls. Section 3.1 establishes the formal definition; Section 3.2 defines the five atomic meta-operations; Section 3.3 proves expressive completeness and Section 3.4 details the data synthesis and training pipeline.
3
Input Tools
Managed Context User query
Meta tools
Compressed Step 1 to 4: Searched for [keyword], visited [source], found key info and concluded that … Step 5: A Google search for [keyword] found 10 results: [#1 Google search result] [#2 Google search result] … Deleted Step 6: [Step deleted] Rollbacked to Step 7: Revert to Step 7 because previous path led to insufficient info …
LongSeeker
Skip
Compress
Rollback
Keep context
Summarize step(s)
Revert to previous step
Snippet
Delete
Keep key info
Remove step
Standard tools
Output Think:
Reasoning path
Meta tool call:
Call & params
Motivation:
Why this helps
Standard tool call:
Call & params
Tool response:
Search & visit results
Search Web
Snippeted Step 8: [#2 Google search result]
Visit Web
Step 9: The info in [source] confirms …
Update for next step
Figure 2: Overview of the Context-ReAct paradigm. Unlike standard ReAct, which passively accumulates history, and unlike prior proactive curation methods (Ye et al., 2025; Yao et al., 2026) that operate at a coarse granularity, Context-ReAct introduces a complete and fine-grained meta-action layer. At each step, the agent co-generates meta-operations (S KIP, C OMPRESS, ROLLBACK, S NIPPET, D ELETE) alongside standard tool calls, enabling elastic context orchestration that spans the full spectrum from lossless extraction to structural backtracking.
3.1
Agentic Paradigm
In this section, we introduction our proposed Context-ReAct with a formal definition. Standard ReAct. In the standard ReAct paradigm (Yao et al., 2023), each step Sistd is defined as
Sistd = (ri , ci , oi ),
(1)
where ri is the chain-of-thought reasoning trace, ci is the tool call, and oi is the environment observation returned by the tool. The context history at time t is the concatenation Ht = [S1std , . . . , Ststd ], which grows monotonically and without bound under this append-only design. As irrelevant observations accumulate, the signal-to-noise ratio of Ht degrades, and |Ht | eventually risks exceeding the model’s context limit. Context-ReAct. To preserve the generality of the standard ReAct paradigm while equipping the agent with dynamic context management, we augment each step by inserting several meta operations between the reasoning trace and the standard tool call. The resulting step structure is
Simeta = (ri , Mi , ci , oi ), (i)
(i)
(2)
(i)
where Mi = [op 1 , op 2 , . . . , op k ] is a list of meta-operations generated by the agent to transform the context before the next step begins. Formally, the effective context Ht′ used at step t+1 is
Ht′ = T (Ht , Mt ),
(3)
where T is the composition of the individual operations in Mt , each drawn from the primitive set (defined in Section 3.2). This mechanism enables the agent to maintain a compact and relevant working memory (|Ht′ | ≪ |Ht | in typical long-horizon tasks) without any external trigger or architectural modification. The design principle is that meta-operations are co-generated with the reasoning trace and tool call in a single, end-to-end generation step, rather than being triggered by an external heuristic such as a length threshold. The agent therefore learns when, where, and how to intervene in its own context as an integral part of its policy; see Figure 2 (right) for illustration. 4
3.2
Atomic Meta-Operations
To enable flexible and task-aware management of the working context, Context-ReAct equips the agent with a set of atomic meta-operations. These operations define the full set of primitive actions that can be applied to the context at each reasoning step. Throughout, let H = [S1 , . . . , Sn ] denote the current history. (1) S KIP is the identity operator. That is, the agent takes no action on the context: S KIP(H) = H,
(4)
which is issued when the current context is already compact, incurring zero additional generation overhead. (2) C OMPRESS performs abstractive summarization over any contiguous range of steps [a, b] (a ≤ b), replacing the context with a summarized step Sa:b , where Σ is the summarized string: C OMPRESS(H, a, b, Σ) = [S1 , . . . , Sa−1 , Sa:b = Σ, Sb+1 , . . . , Sn ].
(5)
Crucially, [a, b] need not be a recent window (Ye et al., 2025): the agent can retroactively recognize that steps from early in the trajectory have become compressible. For example, earlier searches may no longer need to be preserved in full once their useful information has been captured by later evidence, allowing the agent to compress them on the fly. This look-back flexibility is unavailable in sliding-window approaches, which can only truncate from one end of the history. (3) ROLLBACK reverts the context to step k by discarding all subsequent steps Sk , . . . , Sn and appending a summarized step that records the reason for backtracking and any transferable insight: ROLLBACK(H, k, Σ) = [S1 , . . . , Sk = Σ].
(6)
ROLLBACK models the structural intuition of branch abandonment in tree-based search (DFS/MCTS) (Shi et al., 2025): when the agent recognizes that a reasoning path has reached a dead end, it discards the failed sub-trajectory while preserving its causal explanation, preventing the same mistake from being repeated. (4) S NIPPET replaces the observation ok of the k -th step with the verbatim substring delimited by the anchor strings pre and suf : S NIPPET(H, k, pre, suf) = S1 , . . . , (rk , ck , ok [pre:suf]), . . . , Sn . (7) Unlike generative summarization, S NIPPET is lossless with respect to the retained segment: it performs pointer-based substring extraction rather than token regeneration, saving token cost and preventing hallucination of precise numerical values, entity names, URLs, or code that must be carried forward exactly. (5) D ELETE removes the k -th step, discarding its reasoning trace, tool call, and observation: D ELETE(H, k) = S1 , . . . , Sk−1 , Sk+1 , . . . , Sn .
(8)
This operation is appropriate when an entire step is uninformative and leaves no useful trace—e.g., a failed or redundant query whose result, reasoning, and call all warrant complete removal to reduce noise. Composite application. Multiple meta-operations may be composed within a single step by listing them sequentially in Mi . This compositionality allows LongSeeker to, for instance, D ELETE a noisy interaction step and simultaneously C OMPRESS a longer historical segment in one step. Figure 3 illustrates the combined effect of these meta-operations on a live trajectory (see Appendix A for a complete example). Rather than receiving the full raw trajectory, the model sees a compact, curated view of its history. As shown in the left half of the figure, Steps 1–4 have been consolidated via C OMPRESS into a single summary sentence that preserves the essential findings while discarding verbose intermediate observations. Step 5 remains unchanged via S KIP. Step 6 has been D ELETEd to eliminate noise from a redundant query. Most significantly, the trajectory has been ROLLBACKed to Step 7, discarding an unproductive sub-chain of exploration. The resulting managed context is minimal yet information-dense, ensuring the model can focus on the remaining open sub-questions. 5
Managed Context ### Question
Find a brand that: • Uses a clinically supported element to solve customer issues • Was acquired by a company whose founder graduated in the 1900s • Has a co-founder from a university that recently announced a merger • Sold products in 30+ countries by 2023
### Previous Steps Compressed Step 1 to 4: Steps 1–3 focused on the cofounder education clue and L'Oréal as a likely acquirer, but found no co-founders from City or St George's universities. Step 4 revealed that L'Oréal acquired Medik8—a clinically supported skincare brand available in over 30 countries. Step 5: Motivation: We have confirmed L'Oréal as the acquirer … Standard tool call: {"name": "search_web", "arguments”: …} Tool response: [Web search results, omitted due to length] Deleted Step 6: [Step deleted] Rollbacked to Step 7: Rollbacked to step 7 because the current approach is not yielding results and produces useless information. The searches for co-founders …
Output <think> We've confirmed CeraVe is in 30+ countries. Still need to verify: 1) Saint Joseph's University's recent merger with Pennsylvania College of Health Sciences, 2) Tom Allison's graduation from there. Let's search for reliable news articles. Also double-check: CeraVe was acquired by L'Oréal in 2017; founder Eugène Schueller graduated in 1904 (matches the 1900s clue). The clinically supported element is ceramides. </think> <meta_tool_call> [{"name": "skip", "arguments": {}}] </meta_tool_call> <motivation> We've verified CeraVe is in 30+ countries. Need to confirm the remaining two clues: Saint Joseph's University's recent merger and Tom Allison's graduation from there. We'll perform targeted web searches to gather evidence. </motivation> <standard_tool_call> {"name": "search_web", "arguments": {"query": ["Saint Joseph's University merger with Pennsylvania College of Health Sciences", "Tom Allison CeraVe co-founder Saint Joseph's University", "CeraVe co-founder Tom Allison education"]}} </standard_tool_call>
Figure 3: Managed context and structured output at a single Context-ReAct step. Left: The curated context after applying meta-operations to the raw trajectory. Steps 1–4 are consolidated via C OMPRESS into a summary preserving essential findings. Step 5 remains unchanged via S KIP. Step 6 is D ELETEd to eliminate noise from a redundant query. The trajectory is ROLLBACKed to Step 7, discarding unproductive exploration. The resulting context is minimal yet information-dense. Right: The four-field structured output containing reasoning, meta-operations, motivation, and the standard tool call.
3.3
Expressive Completeness and Principled Redundancy
We now analyze their theoretical properties with a simple d ceduction to justify both the completeness and practical utility of the Context-ReAct action set. Theorem 3.1 (Expressive Completeness). The meta-action set O = {S KIP, C OMPRESS, ROLLBACK, S NIPPET, D ELETE} is expressively complete: for any Hin , Htarget ∈ H, there exists a finite sequence of operations from O that transforms Hin into Htarget . Proof. It suffices to show that C OMPRESS alone is a universal string rewriting operator over H. By definition, C OMPRESS(H, 1, |H|, Σ) replaces the entire history H with an arbitrary string Σ ∈ V ∗ . Setting Σ = Htarget yields H ′ = Htarget in a single operation. Since a single element of O can reach any target from any source, the full set O is trivially complete. Although C OMPRESS alone is sufficient for theoretical completeness, the remaining four operators provide practical structure by guiding the agent to manage context more efficiently and reliably. Each operator addresses a distinct operational need: • S KIP: identity. S KIP(H) ≡ C OMPRESS(H, 1, |H|, H) indicates preservation of the entire history. • ROLLBACK: structural search prior. ROLLBACK(H, k, Σ) ≡ C OMPRESS(H, k, |H|, Σ), but framing it as “rollback to step k ” gives the model a clearer inductive bias: it should discard incorrect reasoning branches, similar to backtracking in tree search. This helps the agent learn when to abandon a failed path, rather than treating it as a generic compression.
6
• S NIPPET: extraction. S NIPPET(H, k, pre, suf) ≡ C OMPRESS(H, k, k, ok [pre:suf]), but generative compression is stochastic and lossy. S NIPPET guarantees exact retention of critical content via pointerbased extraction, which is critical when the retained segment contains numerical values, entity names, or code that must be reproduced verbatim in later reasoning steps. • D ELETE: complete step removal. D ELETE(H, k) ≡ C OMPRESS(H, k, k, ∅), but framing it as “delete step k ” makes the operation explicit: the entire step k is removed from the context. In summary, the five atomic meta-operations can be interpreted as specialized transformations over distinct subspaces of the space of all possible context. By partitioning the context transformation space in this way, the agent can maintain a compact, relevant, and reliable working context throughout long-horizon reasoning. This structured decomposition also aligns with the Minimum Description Length principle (Rissanen, 1978; Grunwald, 2004), providing a rationale for why these specialized operators are beneficial in practice.
3.4
Data Synthesis and Training
Trajectory synthesis. Training LongSeeker via Context-ReAct requires trajectories that contain not only correct final answers but also high-quality context management decisions at intermediate steps—a supervision signal absent from all existing datasets. We construct a corpus of 10,000 annotated trajectories through a two-stage pipeline. Stage 1: Seed question collection. We sample 10,000 complex, multi-hop questions from OpenSeeker (Du et al., 2026b,a), comprising 9,000 English and 1,000 Chinese questions, filtering for questions that require substantive multi-step reasoning to answer. Stage 2: Context-ReAct trajectory rollout. Each question is solved by DeepSeek V3.2 (DeepSeek-AI, 2025) acting as the teacher model and operating under the full Context-ReAct paradigm. At every step, the teacher directly generates the complete four-field structured output—<think>, <meta_tool_call>, <motivation>, and <standard_tool_call>—in a single pass, producing context management decisions and the next tool call jointly. Trajectories with correct format constitute the final training set. Supervised fine-tuning. We fine-tune Qwen 3 30B-A3B (Yang et al., 2025) on the annotated corpus via standard next-token prediction:
LSFT = −
T X X t=1
(t) (t) ′ log pθ xj x<j , Ht−1 ,
(9)
j
where x(t) is the full structured output at step t—including the chain-of-thought, meta-tool call, motivation, ′ and standard tool call—and Ht−1 is the context after applying the meta-operations from the previous step. Computing the loss over the entire structured output forces the model to jointly learn which meta-operation to invoke, when to invoke it, and how to use the standard tool given the current context.
4
Experiments
4.1
Experimental Setup
Evaluations. We evaluate our LongSeeker on four key benchmarks spanning targeted information-seeking and general agent capabilities. BrowseComp (Wei et al., 2025a) and BrowseComp-ZH (Zhou et al., 2025b) evaluate multi-step navigation and hard information retrieval in English and Chinese, respectively (sampling 200 questions from each benchmark due to resource constraints). xbench (Chen et al., 2025) assesses complex deep research capabilities including planning, reasoning, and synthesis across profession-aligned real-world tasks. Finally, GAIA (Mialon et al., 2023) (text-only subset) evaluates general agent capabilities requiring combined web browsing, tool use, and multi-step reasoning. We set the max tool call as 300 for 7
Table 1: Main results. LongSeeker, trained under the Context-ReAct paradigm, outperforms GPT-5 and Gemini-3.0-Pro on BrowseComp despite having only 30B parameters, highlighting the effectiveness and potential of Context-ReAct. Scores marked with ∗ denote ReAct-based agents without context management on BrowseComp and BrowseComp-ZH, while “–” indicates unavailable or unknown results. Model
Param
Training
BrowseComp
BrowseComp-ZH
xbench-2505
GAIA-text
63.0 66.8 62.4∗ 82.4 65.0∗ 66.6∗
77.8 78.0 72.0
76.4 74.8 71.5 78.6 75.1 -
Foundation Model with Tools GPT-5 Gemini-3.0-Pro Claude-Opus-4.5 Seed-2.0-Pro DeepSeek-V3.2 GLM-4.7
671B 358B
-
54.9 59.2 67.8 77.3 67.6 67.5 Search Agent
MiroThinker-1.7-mini MiroThinker-1.5-mini RedSearcher IterResearch AgentFold Tongyi-DeepResearch OpenSeeker
30B 30B 30B 30B 30B 30B 30B
CPT + SFT + RL CPT + SFT + RL CPT + SFT + RL SFT + RL SFT CPT + SFT + RL SFT
67.9 56.1 57.4 37.3 36.2 43.4∗ 29.5∗
72.3 66.8 58.2 45.2 47.3 46.7∗ 48.4∗
73.1 71.0 75.0 74.0
80.3 72.0 80.1 72.8 67.0 70.9 -
LongSeeker
30B
SFT
61.5
62.5
78.0
77.7
all benchmarks. For BrowseComp and BrowseComp-ZH, we also apply the discard-all technique and allow for 5 rounds at maximum following MiroThinker (Team et al., 2026). Baselines. To validate the effectiveness of LongSeeker, we compare it against several state-of-the-art systems categorized into two groups: (1) foundation models with tools, comprising frontier proprietary systems such as GPT-5 (OpenAI, 2025b), Gemini-3.0-Pro (Google DeepMind, 2025), Claude-Opus4.5 (Anthropic, 2025), and Seed-2.0-Pro (ByteDance Seed Team, 2026), alongside open-weight models DeepSeek-V3.2 (DeepSeek-AI, 2025) and GLM-4.7 (Zhipu AI, 2025); and (2) search agents, which serve as direct, comparable-scale benchmarks at 30B parameters, including MiroThinker series (Team et al., 2026), REDSearcher (Zheng et al., 2026), IterResearch (Chen et al., 2026), AgentFold (Ye et al., 2025), Tongyi-DeepResearch (Team et al., 2025), and OpenSeeker (Du et al., 2026b). This diverse baseline set covers representative paradigms in contemporary agentic search. All baseline results are sourced from official publications or publicly available evaluation platforms.
4.2
Results and Analysis
Main results. Table 1 presents the primary evaluation results on BrowseComp and BrowseComp-ZH. LongSeeker achieves 61.5 on BrowseComp and 62.5 on BrowseComp-ZH, establishing strong performance among 30B-scale open-source search agents. Notably, LongSeeker exceeds MiroThinker-1.5-mini (56.1), Tongyi-DeepResearch (43.4), IterResearch (37.3), AgentFold (36.2), and OpenSeeker-v1 (29.5). Extending evaluation to xbench and GAIA, LongSeeker achieves 78.0 on xbench-2505 and 77.7 on GAIA-text. These scores confirm that the benefits of Elastic Context Orchestration generalize beyond purely informationseeking tasks to broader agent capabilities, with LongSeeker establishing competitive performance across diverse benchmark suites. Context growth dynamics. To empirically validate the efficacy of our context management paradigm, we trace the trajectory length and corresponding context token count across 200 questions sampled from BrowseComp. As depicted in Figure 4a, we plot the surviving trajectories at each turn alongside their 8
(a) Context Growth Dynamics of LongSeeker
(b) Meta-operation distribution of LongSeeker
Figure 4: Analysis of LongSeeker’s context management on 200 trajectories sampled from BrowseComp. (a) The average context token count remains stable and well bounded (plateauing around 15k tokens) over long horizons, in contrast to the explosive linear growth of ReAct-based DeepSeek-V3.2. The managed context is highly compact, utilizing a mere fraction of the LongSeeker’s 256k capacity. (b) LongSeeker learns through training to utilize all five meta-operations and effectively invoke and compose them to solve long-horizon tasks. average accumulated context tokens. Unlike ReAct-based DeepSeek-V3.2, which suffers from rapid, unbounded context inflation as observations are passively appended step-by-step, LongSeeker maintains a remarkably stable and concise working memory. The token count initially scales with the problem depth but soon reaches a plateau, staying under 15k tokens even at extended horizons of 300 steps. This stabilized growth is a direct consequence of our complete and fine-grained meta-operations: rather than accumulating noise, the model learns to dynamically purge failed branches (ROLLBACK), discard irrelevant retrievals (D ELETE), extract only essential snippets (S NIPPET), and abstract verbose history (C OMPRESS). Consequently, the context remains highly compact and information-dense. LongSeeker delivers competitive performance on long-horizon benchmarks while utilizing only a fraction of the underlying model’s maximum 256k context window. This vast remaining capacity leaves ample headroom for tackling longer and more complex exploratory tasks. This confirms that the model has internalized how to efficiently deploy our atomic meta tools, retaining critical reasoning signals while minimizing distracting noise. Meta-operation usage. Figure 4b shows the usage distribution of the five meta-operations of LongSeeker. We observe that LongSeeker effectively leverages the full set of atomic operations—Skip, Compress, Rollback, Snippet, and Delete—across trajectories, suggesting that it has acquired a robust strategy for invoking and composing meta-operation to handle complex long-horizon tasks. This ability to coherently and purposefully navigate the context-manipulation space contributes to its strong overall performance. We also observe a mild imbalance, where Snippet and Delete are used less frequently. This likely stems from the nature of long-horizon search: early in the process, it is difficult to confidently identify irrelevant information, so LongSeeker tends to preserve more context and adopts a conservative pruning strategy. Comparison of Context Management Strategies. To evaluate the effectiveness of the Context-ReAct paradigm, we conduct controlled experiments on BrowseComp under a unified setup, where all methods are built upon the same base model (DeepSeek-V3.2) and share identical configurations. We compare ContextReAct with two commonly used context management strategies used in DeepSeek-V3.2 (DeepSeek-AI, 2025): (1) Summary, which compresses the overflowed trajectory into a summary and resumes the rollout from the condensed context; and (2) Discard-all, which resets the context by removing all previous tool-call history, similar to reinitializing with a fresh context. As shown in Figure 5, Context-ReAct consistently achieves the highest performance under the same step budget, demonstrating its superior effectiveness in
9
Figure 5: Effectiveness of the Context-ReAct paradigm on BrowseComp compared to other context management strategies. Context-ReAct achieves better performance under the same step budget. long-horizon tasks through elastic context control that enables adaptive management and utilization of context throughout extended trajectories.
5
Conclusion
We propose Context-ReAct, a general agentic paradigm for elastic context orchestration that enables agents to jointly generate reasoning, context meta-operations, and tool calls at each step. Context-ReAct defines five atomic operations—Skip, Compress, Rollback, Snippet and Delete. Together, these operations provide a complete and fine-grained mechanism for multi-resolution control over the evolving working context. We train LongSeeker-30B based on this paradigm and demonstrate competitive performance on long-horizon search benchmarks, notably surpassing Tongyi DeepResearch and AgentFold on BrowseComp. Our experiments empirically validate that the application of these atomic tools yields significantly more compact and efficient context management compared to append-only or coarse-grained truncation approaches, enabling sustained high performance at long horizons with strong potential for harder tasks. Future work. Our current implementation leverages SFT on synthesized trajectories without rejection sampling or advanced exploration strategies. One direction involves applying RL to optimize metaoperation usage, enabling agents to explore the action spaces. Furthermore, we envision Context-ReAct as a foundational architectural paradigm rather than a search-specific solution. Its core philosophy of state-dependent fidelity is inherently domain-agnostic, offering a scalable blueprint for other long-horizon challenges such as autonomous software engineering, large-scale legal discovery, and multi-modal scientific reasoning, where the ability to fluidly restructure massive working contexts is critical.
References Anthropic. Claude opus 4.5 system card, 2025. claude-opus-4-5-system-card.
URL https://www.anthropic.com/
ByteDance Seed Team. Seed2.0 model card: Towards intelligence frontier for real-world complexity, 2026. URL https://lf3-static.bytednsdoc.com/obj/eden-cn/lapzild-tss/ ljhwZthlaukjlkulzlp/seed2/0214/Seed2.0%20Model%20Card.pdf. Guoxin Chen, Zile Qiao, Xuanzhong Chen, Donglei Yu, Haotian Xu, Wayne Xin Zhao, Ruihua Song, Wenbiao Yin, Huifeng Yin, Liwen Zhang, Kuan Li, Minpeng Liao, Yong Jiang, Pengjun Xie, Fei Huang, and Jingren Zhou. Iterresearch: Rethinking long-horizon agents with interaction scaling. arXiv preprint arXiv:2511.07327, 2026. 10
Kaiyuan Chen, Yixin Ren, Yang Liu, Xiaobo Hu, Haotong Tian, Tianbao Xie, Fangfu Liu, Haoye Zhang, Hongzhang Liu, Yuan Gong, et al. xbench: Tracking agents productivity scaling with profession-aligned real-world evaluations. arXiv preprint arXiv:2506.13651, 2025. DeepSeek-AI. Deepseek-v3.2: Pushing the frontier of open large language models, 2025. URL https: //arxiv.org/abs/2512.02556. Yuwen Du, Rui Ye, Shuo Tang, Keduan Huang, Xinyu Zhu, Yuzhu Cai, and Siheng Chen. Openseeker-v2: Pushing the limits of search agents with informative and high-difficulty trajectories. arXiv preprint arXiv:2605.04036, 2026a. Yuwen Du, Rui Ye, Shuo Tang, Xinyu Zhu, Yijun Lu, Yuzhu Cai, and Siheng Chen. Openseeker: Democratizing frontier search agents by fully open-sourcing training data. arXiv preprint arXiv:2603.15594, 2026b. Google DeepMind. Model evaluation - approach, methodology & results, gemini 3 pro, 2025. URL https://storage.googleapis.com/deepmind-media/gemini/gemini_ 3_pro_model_evaluation.pdf. Peter Grunwald. A tutorial introduction to the minimum description length principle, 2004. URL http: //arxiv.org/abs/math/0406077. Kuan Li, Zhongwang Zhang, Huifeng Yin, Liwen Zhang, Litu Ou, Jialong Wu, Wenbiao Yin, Baixuan Li, Zhengwei Tao, Xinyu Wang, Weizhou Shen, Junkai Zhang, Dingchu Zhang, Xixi Wu, Yong Jiang, Ming Yan, Pengjun Xie, Fei Huang, and Jingren Zhou. Websailor: Navigating super-human reasoning for web agent, 2025. URL https://arxiv.org/abs/2507.02592. Miao Lu, Weiwei Sun, Weihua Du, Zhan Ling, Xuesong Yao, Kang Liu, and Jiecao Chen. Scaling llm multi-turn rl with end-to-end summarization-based context management, 2025. URL https: //arxiv.org/abs/2510.06727. Grégoire Mialon, Clémentine Fourrier, Craig Swift, Thomas Wolf, Yann LeCun, and Thomas Scialom. Gaia: a benchmark for general ai assistants, 2023. OpenAI. Deep research system card, deep-research-system-card.pdf.
2025.
OpenAI. Deep research system card, February 2025a. deep-research-system-card.pdf.
URL https://cdn.openai.com/ URL https://cdn.openai.com/
OpenAI. Gpt-5 system card, 2025b. URL https://arxiv.org/abs/2601.03267. Jorma Rissanen. Modeling by shortest data description. Automatica, 14(5):465–471, 1978. doi: 10.1016/ 0005-1098(78)90005-5. Zijing Shi, Meng Fang, and Ling Chen. Monte carlo planning with large language model for text-based game agents, 2025. URL https://arxiv.org/abs/2504.16855. MiroMind Team, S. Bai, L. Bing, L. Lei, R. Li, X. Li, X. Lin, E. Min, L. Su, B. Wang, L. Wang, L. Wang, S. Wang, X. Wang, Y. Zhang, Z. Zhang, G. Chen, L. Chen, Z. Cheng, Y. Deng, Z. Huang, D. Ng, J. Ni, Q. Ren, X. Tang, B.L. Wang, H. Wang, N. Wang, C. Wei, Q. Wu, J. Xia, Y. Xiao, H. Xu, X. Xu, C. Xue, Z. Yang, Z. Yang, F. Ye, H. Ye, J. Yu, C. Zhang, W. Zhang, H. Zhao, and P. Zhu. Mirothinker-1.7 & h1: Towards heavy-duty research agents via verification, 2026. URL https://arxiv.org/abs/2603.15726.
11
Tongyi DeepResearch Team, Baixuan Li, Bo Zhang, Dingchu Zhang, Fei Huang, Guangyu Li, Guoxin Chen, Huifeng Yin, Jialong Wu, Jingren Zhou, Kuan Li, Liangcai Su, Litu Ou, Liwen Zhang, Pengjun Xie, Rui Ye, Wenbiao Yin, Xinmiao Yu, Xinyu Wang, Xixi Wu, Xuanzhong Chen, Yida Zhao, Zhen Zhang, Zhengwei Tao, Zhongwang Zhang, Zile Qiao, Chenxi Wang, Donglei Yu, Gang Fu, Haiyang Shen, Jiayin Yang, Jun Lin, Junkai Zhang, Kui Zeng, Li Yang, Hailong Yin, Maojia Song, Ming Yan, Minpeng Liao, Peng Xia, Qian Xiao, Rui Min, Ruixue Ding, Runnan Fang, Shaowei Chen, Shen Huang, Shihang Wang, Shihao Cai, Weizhou Shen, Xiaobin Wang, Xin Guan, Xinyu Geng, Yingcheng Shi, Yuning Wu, Zhuo Chen, Zijian Li, and Yong Jiang. Tongyi deepresearch technical report, 2025. URL https://arxiv.org/abs/2510.24701. Jason Wei, Zhiqing Sun, Spencer Papay, Scott McKinney, Jeffrey Han, Isa Fulford, Hyung Won Chung, Alex Tachard Passos, William Fedus, and Amelia Glaese. Browsecomp: A simple yet challenging benchmark for browsing agents. arXiv preprint arXiv:2504.12516, 2025a. Jason Wei, Zhiqing Sun, Spencer Papay, Scott McKinney, Jeffrey Han, Isa Fulford, Hyung Won Chung, Alex Tachard Passos, William Fedus, and Amelia Glaese. Browsecomp: A simple yet challenging benchmark for browsing agents, 2025b. URL https://arxiv.org/abs/2504.12516. An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, Chujie Zheng, Dayiheng Liu, Fan Zhou, Fei Huang, Feng Hu, Hao Ge, Haoran Wei, Huan Lin, Jialong Tang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jing Zhou, Jingren Zhou, Junyang Lin, Kai Dang, Keqin Bao, Kexin Yang, Le Yu, Lianghao Deng, Mei Li, Mingfeng Xue, Mingze Li, Pei Zhang, Peng Wang, Qin Zhu, Rui Men, Ruize Gao, Shixuan Liu, Shuang Luo, Tianhao Li, Tianyi Tang, Wenbiao Yin, Xingzhang Ren, Xinyu Wang, Xinyu Zhang, Xuancheng Ren, Yang Fan, Yang Su, Yichang Zhang, Yinger Zhang, Yu Wan, Yuqiong Liu, Zekun Wang, Zeyu Cui, Zhenru Zhang, Zhipeng Zhou, and Zihan Qiu. Qwen3 technical report, 2025. URL https://arxiv.org/abs/2505.09388. Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. React: Synergizing reasoning and acting in language models, 2023. URL https://arxiv.org/abs/ 2210.03629. Yilun Yao, Shan Huang, Elsie Dai, Zhewen Tan, Zhenyu Duan, Shousheng Jia, Yanbing Jiang, and Tong Yang. Arc: Active and reflection-driven context management for long-horizon information seeking agents, 2026. URL https://arxiv.org/abs/2601.12030. Rui Ye, Zhongwang Zhang, Kuan Li, Huifeng Yin, Zhengwei Tao, Yida Zhao, Liangcai Su, Liwen Zhang, Zile Qiao, Xinyu Wang, Pengjun Xie, Fei Huang, Siheng Chen, Jingren Zhou, and Yong Jiang. Agentfold: Long-horizon web agents with proactive context management, 2025. URL https: //arxiv.org/abs/2510.24699. Hongli Yu, Tinghong Chen, Jiangtao Feng, Jiangjie Chen, Weinan Dai, Qiying Yu, Ya-Qin Zhang, WeiYing Ma, Jingjing Liu, Mingxuan Wang, and Hao Zhou. Memagent: Reshaping long-context llm with multi-conv rl-based memory agent, 2025. URL https://arxiv.org/abs/2507.02259. Chu Zheng, Xiao Wang, Jack Hong, Huiming Fan, Yuqi Huang, Yue Yang, Guohai Xu, Chenxiao Zhao, Cheng Xiang, Shengchao Hu, Dongdong Kuang, Ming Liu, Bing Qin, and Xing Yu. Redsearcher: A scalable and cost-efficient framework for long-horizon search agents. arXiv preprint arXiv:2602.14234, 2026. Zhipu AI. Glm-4.7: Advancing the coding capability, December 2025. URL https://z.ai/blog/ glm-4.7.
12
Peilin Zhou, Bruce Leon, Xiang Ying, Can Zhang, Yifan Shao, Qichen Ye, Dading Chong, Zhiling Jin, Chenxuan Xie, Meng Cao, Yuxin Gu, Sixin Hong, Jing Ren, Jian Chen, Chao Liu, and Yining Hua. Browsecomp-zh: Benchmarking web browsing ability of large language models in chinese, 2025a. URL https://arxiv.org/abs/2504.19314. Peilin Zhou, Bruce Leon, Xiang Ying, Can Zhang, Yifan Shao, Qichen Ye, Dading Chong, Zhiling Jin, Chenxuan Xie, Meng Cao, et al. Browsecomp-zh: Benchmarking web browsing ability of large language models in chinese. arXiv preprint arXiv:2504.19314, 2025b. Zijian Zhou, Ao Qu, Zhaoxuan Wu, Sunghwan Kim, Alok Prakash, Daniela Rus, Jinhua Zhao, Bryan Kian Hsiang Low, and Paul Pu Liang. Mem1: Learning to synergize memory and reasoning for efficient long-horizon agents, 2025c. URL https://arxiv.org/abs/2506.15841.
13
A
Case Study
Figure 6 shows the managed context after applying meta-operations, and Figure 7 shows the corresponding structured output from LongSeeker. Managed Context
### Question Give me the first and last name of the first Prime Minister of this country by identifying the country based on the following clues, as of 2023: • it is a landlocked country and two of its neighbors are also landlocked countries • its literacy rate in 2022 increased by 1.16% as compared to 2014 • it is home to a species from the Family Chamaeleonidae named after the individual who collected its holotype. ### Previous Steps [Compressed Step 1~5] We performed initial searches to understand the clues. We identified that there are 44 landlocked countries, and many have at least two landlocked neighbors. We searched for chameleon species named after their holotype collector and found examples like Trioceros jacksonii (Jackson's chameleon) named after F.J. Jackson, but this species is from Kenya and Tanzania (not landlocked). We also searched for literacy rate increase of 1.16% between 2014 and 2022, but did not find a specific country yet. We need to narrow down by finding a landlocked country that hosts a chameleon species named after its holotype collector and has a literacy rate increase of exactly 1.16% from 2014 to 2022. [Rollback to Step 6] The previous steps identified Uganda as a candidate based on the chameleon clue, but the literacy rate increase does not match (12.08% vs 1.16%). We must consider other landlocked countries with at least two landlocked neighbors and verify the chameleon species and literacy rate clues more carefully. We also need to ensure we are using the correct literacy rate metric (possibly youth literacy or adult literacy from a specific source). We will start by listing all landlocked countries, then filter by those with at least two landlocked neighbors, then search for chameleon species named after their holotype collector in those countries, and finally check literacy rate changes. [Compressed Step 7~13] We attempted to find the country by searching for the precise literacy rate increase of 1.16% from 2014 to 2022, but the search results did not directly yield a country. We also searched for chameleon species named after their holotype collector, finding Jackson's chameleon (Trioceros jacksonii) named after F.J. Jackson, but it is found in Kenya and Tanzania (not landlocked). We also considered Johnston's chameleon (Trioceros johnstoni) but need to confirm if it is named after the holotype collector. We then listed landlocked African countries with at least two landlocked neighbors: Botswana, Burkina Faso, Central African Republic, Chad, Mali, Niger, Rwanda, South Sudan, Uganda, Zambia, Zimbabwe. We need to verify the literacy rate increase and chameleon clue for each candidate. [Deleted Step 14] [Step deleted] [Compressed Step 15] We searched for literacy rate data (2014 and 2022) for each candidate landlocked African country with at least two landlocked neighbors. The results show that for most countries, the increase from 2014 to 2022 is not 1.16%. However, Zimbabwe's literacy rate increased from 88.7% in 2014 to 89.85% in 2022, a rise of 1.15% (very close to 1.16%, possibly rounding). Other candidates like Uganda, Rwanda, Zambia, etc., show increases far above 1.16%. Thus, Zimbabwe emerges as a strong candidate based on the literacy rate clue. [Snippeted Step 16] **Motivation** We have identified Zimbabwe as a candidate based on its ~1.16% literacy rate increase. We have already confirmed it's landlocked with at least two landlocked neighbors (Zambia, Botswana). The remaining clue is a Chamaeleonidae species found in Zimbabwe and named after its holotype collector. We will search for that species, then identify Zimbabwe's first Prime Minister (likely postindependence). **Standard tool call** {"name": "search_web", "arguments": {"query": ["chameleon species Zimbabwe named after collector", "Zimbabwe chameleon holotype collector", "Chamaeleonidae Zimbabwe named after", "chameleon named after collector found in Zimbabwe"]}} **Tool response** 3. [Marshall's pygmy chameleon – Wikipedia] The specific name, marshalli, is in honor of British entomologist Guy Anstruther Knox Marshall, who collected the holotype. ... Zimbabwe and the adjacent upland ... 4. [Chamaeleo dilepis - The Reptile Database - Reptarium] Taxonomic database that provides basic information about all living reptile species, such as turtles, snakes, lizards, and crocodiles, as well as tuataras ...
Figure 6: Complete case study showing managed context at a reasoning step. The trajectory demonstrates the combined effect of C OMPRESS, ROLLBACK, D ELETE, and S NIPPET operations.
14
Output <think> We have identified Zimbabwe as a candidate based on literacy rate increase (approx 1.15-1.16%). Now we need to verify the chameleon clue: "it is home to a species from the Family Chamaeleonidae named after the individual who collected its holotype." We need to find a chameleon species found in Zimbabwe that is named after the person who collected its holotype. The search results mention "Marshall's pygmy chameleon" (Rhampholeon marshalli) named after Guy Anstruther Knox Marshall, who collected the holotype. However, we need to confirm that this species is found in Zimbabwe. The Wikipedia page says it is found in Zimbabwe and Mozambique. That seems promising. But we should verify that the species is indeed from the Family Chamaeleonidae (it is a chameleon). Also, we need to ensure that the naming is indeed after the collector of the holotype. The Wikipedia snippet says "The specific name, marshalli, is in honor of British entomologist Guy Anstruther Knox Marshall, who collected the holotype." That matches the clue. We also need to confirm the literacy rate increase exactly. In step 14, we found Zimbabwe's literacy rate increased from 88.7% in 2014 to 89.85% in 2022, which is a 1.15% increase. The clue says 1.16% increase. Possibly rounding difference. We should verify the exact numbers from a reliable source. Additionally, we need to identify the first Prime Minister of Zimbabwe. The question says \"the first Prime Minister of this country by identifying the country based on the following clues, as of 2023.\" That likely means the first Prime Minister after independence (or the first in the country's history). Zimbabwe's first Prime Minister was Robert Mugabe (took office 1980). However, there is also the position of Prime Minister of Southern Rhodesia (before independence). But the question says \"as of 2023\", meaning we are considering the country as it is now. The first Prime Minister of independent Zimbabwe is Robert Mugabe. But we need to confirm. We should also verify that Zimbabwe is indeed landlocked and has at least two landlocked neighbors (Zambia and Botswana). That is correct. Thus, we need to gather confirming evidence for each clue. Let's visit relevant pages. First, visit the Wikipedia page for Marshall's pygmy chameleon to confirm the naming and distribution. Second, visit a reliable source for Zimbabwe literacy rates (e.g., World Bank, CIA World Factbook, or other statistical databases) to confirm the 1.16% increase. Third, confirm the first Prime Minister of Zimbabwe. </think> <meta_tool_call> [{"name": "skip", "arguments": {}}] </meta_tool_call> <motivation> We have identified Zimbabwe as a strong candidate based on the literacy rate increase (approximately 1.15-1.16%) and the fact that it is landlocked with at least two landlocked neighbors. Now we need to verify the remaining clue about the chameleon species named after its holotype collector. The search results suggest Marshall's pygmy chameleon (Rhampholeon marshalli) is named after collector Guy Anstruther Knox Marshall and is found in Zimbabwe. We should visit the Wikipedia page to confirm this. Additionally, we need to verify the literacy rate increase with precise data and confirm the first Prime Minister of Zimbabwe. We'll use visit_web to gather this information. </motivation> <standard_tool_call> {"name": "visit_web", "arguments": {"url”: [omitted due to length], "goal": "1. Confirm that Marshall's pygmy chameleon is named after the collector of its holotype and is found in Zimbabwe. 2. Find literacy rate data for Zimbabwe in 2014 and 2022 to verify 1.16% increase. 3. Identify the first Prime Minister of Zimbabwe."}} </standard_tool_call>
Figure 7: Complete structured output from LongSeeker, including reasoning, meta-tool calls, motivation, and standard tool call.
15