acceptodds
Under review as a conference paper at ICLR 2027

LongSeeker: Elastic Context Orchestration for Long-Horizon Search Agents

Abstract

Long-horizon search agents must manage a rapidly growing working context as they reason, call tools, and observe information. Naively accumulating all interme- diate 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 rel- evance 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. Together, these operations form a fine-grained action space spanning abstractive compression, exact extraction, deletion, structural rollback, and no modification. 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 (43.4% and 46.7%) and AgentFold (36.2% and 47.3%). These re- sults 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.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Search agents, Context management

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_YGhwOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes adding context-management operations to the ReAct loop, which it calls Context-ReAct. Five operations are introduced: Skip, Compress, Snippet, Delete, and Rollback. At each step, the agent generates in one pass its reasoning, a list of operations, a motivation, and a tool call. The operations can be combined and can act on any earlier part of the trajectory. There is no separate summary model or tool; the model writes the Compress and Rollback summaries itself in the same generation pass. Based on this, the authors train LongSeeker, a Qwen3-30B-A3B model fine-tuned on 10k trajectories generated by DeepSeek-V3.2 under this scaffold. The paper reports 61.5% on BrowseComp and 62.5% on BrowseComp-ZH with a discard-all retry wrapper (56.0 / 57.3 without it), plus 78.0 on xbench-2505 and 77.7 on GAIA-text. On BrowseComp this is above Tongyi DeepResearch (43.4), AgentFold (36.2), GPT-5 (54.9) and Gemini-3.0-Pro (59.2), but below MiroThinker-1.7-mini (67.9) and several frontier models. In the matched comparison against a ReAct-SFT model with the same teacher, base model, data and recipe, the gain is 12.5 points on BrowseComp (43.5 → 56.0), 1.3 on BrowseComp-ZH, 3.2 on GAIA-text, and 0 on xbench-2505, with 43.4% fewer tokens but about twice as many tool calls.

Critical strengths
  1. A controlled experiment clearly shows the effect of the proposed method. In Table 2, Context-ReAct-SFT and ReAct-SFT use the same training questions, teacher model, base model and SFT configuration, and only differ in the scaffold. Under this setting, BrowseComp rises from 43.5 to 56.0, BrowseComp-ZH from 56.0 to 57.3, and GAIA-text from 74.5 to 77.7. This is important because it clearly shows the gain comes from the proposed method itself.

  2. The five operations are clear, easy to understand, and precisely defined. Two of them target known failure modes of summary-based curation. Snippet keeps exact strings, such as numbers and URLs, that summaries tend to corrupt, and Rollback lets the agent abandon a dead end while recording why. This is important because it makes the operations easy for other agents to adopt.

Critical weaknesses
  1. Some relatively new and related work are not cited and discussed, such as Sculptor (ICLR 2026; tools to fragment, summarize, hide, restore and search the context), Context-Folding (ICML 2026; branch into a sub-trajectory and fold it into a summary; distinct from the cited AgentFold), which impact the novelty of the paper.

  2. Comparison experiments in Table 1 are not controlled so the conclusion comes with noise.

2.1. A random 400 question subset of BrowseComp, about 32% of 1266 questions, while the baseline number are on the fullset for GPT-5 and Gemini-3.0-Pro, which could fall into the same confidence interval by considering the variance from this subsampling.

2.2. DeepSeek-V3.2 judge with the author's own prompt.

2.3. There is some inconsistency between reporting the model performance for discard-all or without discard-all. discard-all retry wrapper with up to 5 fresh attempts. Table 1 and Table 2 used the same chekpoint, and the only differences are with and without discard all, the whole +5.5 / +5.2 gap on BrowseComp / BrowseComp-ZH comes this discard all and retry. In Section 4.2 the 43.4% token saving comes with about twice as many tool calls (63.65 → 132.50), measured without discard-all. So model performance additional gain also could comes from this discard-all and retry.

Questions for authors
  1. Are the Table 1 baseline numbers measured on the same 400-question BrowseComp subset with the same DeepSeek-V3.2 judge and the same discard-all wrapper? If not, please report confidence intervals for LongSeeker's scores next to the baselines.

  2. Given roughly twice as many tool calls, what are the wall-clock time and cost per trajectory for LongSeeker and for ReAct-SFT? How many discard-all retries did Table 1 actually use on average, and what is the token cost including them?

  3. Do you have any evidence from the trained model on the contribution of individual operations, for example from disabling Snippet or Rollback at inference?

  4. How does Context-ReAct differ from Sculptor, Memory-as-Action and Context-Folding, which also let an agent edit its own context through explicit operations, and does any existing result isolate those differences?

  5. Training trajectories are filtered on format only (Section 3.3), so the student also learns from the teacher's wrong answers. What fraction of the 10k teacher trajectories reach a correct final answer, and does filtering on correctness (or rejection sampling) change the Table 2 result?

5 comments

Thank you for your assessment of our controlled experiments and fine-grained operation design. We appreciate your constructive feedback and address your questions below.

1. [Q1] Evaluation Protocol and Statistical Uncertainty

Table 1 provides a reference for LongSeeker's performance relative to existing systems. Baseline scores follow their respective reported evaluation protocols. For BrowseComp, LongSeeker achieves 61.5% accuracy, with a 95% run-level confidence interval of [60.4%, 62.4%].

Importantly, Table 2 provides a controlled comparison using the same base model, teacher, training questions, SFT configuration, and evaluation protocol. Under this setting, Context-ReAct-SFT improves BrowseComp accuracy from 43.5% to 56.0% (+12.5 pp), providing more direct evidence for the effectiveness of the proposed framework.

2. [Q2] Discard-all Retries and Inference Cost

Following MiroThinker, we adopt discard-all, which resets the interaction history when a rollout reaches the 200-step limit without producing an answer, allowing up to five rounds. On BrowseComp, discard-all incurs an average of 0.89 additional retries per question, increasing total token consumption by 174.5% compared with the setting without discard-all.

To ensure a fair comparison, we additionally evaluate ReAct-SFT under the same discard-all protocol as Context-ReAct-SFT, keeping the base model, training data, and SFT settings unchanged.

MethodBenchmarkw/o Discard-allw/ Discard-all
ReAct-SFTBrowseComp43.550.4
BrowseComp-ZH56.060.9
Context-ReAct-SFTBrowseComp56.061.5
BrowseComp-ZH57.362.5

Context-ReAct consistently outperforms ReAct under both settings. With matched discard-all, it retains gains of 11.1 pp on BrowseComp and 1.6 pp on BrowseComp-ZH, demonstrating that its improvements cannot be attributed solely to additional retries.

Regarding inference efficiency, we extend the token analysis to all 400 BrowseComp questions without discard-all. Context-ReAct reduces total token consumption by 50.4% compared with ReAct-SFT (1,622.8K vs. 3,271.3K tokens per trajectory), despite more tool calls (129.48 vs. 67.91). Model inference dominates computational cost, while Search/Visit API calls are relatively inexpensive. Wall-clock time is affected by API latency fluctuations and concurrent execution. We therefore use total token consumption as the primary measure of inference efficiency.

3. [Q3] Contribution of Individual Operations

To evaluate the contribution of each meta-operation, we conduct ablations on the trained LongSeeker-30B-SFT model using BrowseComp. Each operation is disabled individually while keeping the same checkpoint and evaluation settings.

MethodBrowseComp (%)Δ (pp)Total Tokens (K)Token Reduction (%)
ReAct-SFT43.5−12.53,271.3—
Full Context-ReAct-SFT56.0—1,622.850.4%
w/o Compress46.8−9.22,741.316.2%
w/o Rollback52.2−3.81,982.939.4%
w/o Snippet54.8−1.21,874.542.7%
w/o Delete55.1−0.91,902.841.8%

Compress and Rollback have the largest impact on accuracy, while Snippet and Delete provide smaller gains. Full Context-ReAct achieves the highest accuracy with a 50.4% token reduction relative to ReAct-SFT. These results support the complementary contributions of fine-grained context-management operations to both task performance and token efficiency.

4. [Q4] Comparison with Related Context-Management Methods

We agree that Sculptor, Memory-as-Action, and Context-Folding are relevant related works. Context-ReAct differs in its unified reasoning paradigm and fine-grained operation space. We clarify these distinctions below.

✓: Native support; △: Partial or indirect support; ✗: No corresponding native mechanism.

MethodArbitrary-range CompressLossless SnippetSelective DeleteStructural RollbackComposable Meta OpsJoint Meta Ops + Tool Generation
AgentFold✗✗✗△✗✓
ARC△✗△△✗✗
Sculptor✓△△✗△✗
Context-Folding✗✗✗△✗✗
MemAct△✗✓△△✗
Context-ReAct (Ours)✓✓✓✓✓✓

Our novelty lies in the unified Context-ReAct paradigm rather than any single operation. Unlike AgentFold's suffix-based folding or Context-Folding's predefined branches, Context-ReAct supports editing arbitrary parts of the reasoning history. Compared with ARC's separate context manager, Sculptor's fragment-based tools, and MemAct's memory editing actions, our framework integrates lossless evidence extraction, selective deletion, and structural rollback into a composable operation space.

More importantly, Context-ReAct jointly generates reasoning, multiple meta-operations, and standard tool calls in a single autoregressive pass. These capabilities are learned through SFT, enabling the task agent itself to decide when, where, and how to manage context.

Table 2 provides a controlled comparison with ReAct-SFT under matched training settings, while our additional operation-level ablations examine the contributions of individual operations. These results provide empirical support for Context-ReAct, whereas the capability comparison highlights its methodological distinctions from prior work.

5. [Q5] Training Data Synthesis and Quality

As described in Sec. 3.3, we collect 10K questions from OpenSeeker (9K English and 1K Chinese) and use DeepSeek-V3.2 to generate Context-ReAct trajectories, retaining trajectories with valid output formats. 63% of the 10K teacher trajectories reach a correct final answer. We also evaluated rejection sampling by retaining only correct trajectories, but BrowseComp accuracy dropped from 56.0% to 53.5%. One possible explanation is that trajectories with incorrect final answers still provide valuable supervision for long-horizon exploration and context-management strategies.

Importantly, our controlled comparison in Table 2 uses the same teacher, base model, training questions, and SFT configuration for ReAct and Context-ReAct, isolating the effect of the agent scaffold and its resulting supervision trajectories.

Thank you again for your valuable feedback. We hope these additional experiments and clarifications address your concerns, and we would greatly appreciate it if you could reconsider your rating.

Dear Reviewer YGhw,

Thank you again for your careful review and constructive feedback. As the discussion period approaches its end, we would like to briefly follow up on our rebuttal.

We have provided additional experiments and analyses addressing your five main concerns:

  • Evaluation Protocol and Statistical Uncertainty (Q1): We clarified the evaluation protocols used in Table 1 and reported a 95% confidence interval. We also emphasized the controlled comparison in Table 2, which demonstrates the effectiveness of Context-ReAct under matched settings.
  • Discard-all Retries and Inference Cost (Q2): We clarified the discard-all mechanism and evaluated ReAct-SFT under matched retry settings, confirming that Context-ReAct maintains its advantage. We further quantified the average number of retries and their additional token overhead, alongside a 50.4% reduction in token consumption without discard-all.
  • Contribution of Individual Operations (Q3): We conducted operation-level ablations, demonstrating that Compress and Rollback have the largest impact, while Snippet and Delete provide complementary improvements.
  • Comparison with Related Methods (Q4): We provided a detailed capability comparison with recent context-management approaches, highlighting Context-ReAct's unified, composable operation space and joint generation of reasoning, meta-operations, and tool calls.
  • Training Data Quality (Q5): We analyzed teacher-generated trajectories and conducted correctness-based rejection sampling, showing that filtering for correct answers does not improve performance.

We would appreciate your feedback on whether these additional results and clarifications address your concerns and would be grateful if you could consider increasing your rating.

Thank you for your time and consideration.

The Authors

Reviewer_Yhi8OpenReview ↗
Rating
2
Confidence
4(+2)
Informativeness
1(−1)
Presentation
2
Soundness
1(−1)
Read review
Paper summary

This paper proposes Context-ReAct, a framework for adaptive context management in long-horizon search agents. Rather than passively accumulating the entire interaction history or relying on fixed-granularity summarization, the agent jointly generates reasoning traces, context-management operations, and tool calls. The framework defines five operations: Skip, Compress, Rollback, Snippet, and Delete, enabling the agent to retain the original context, summarize historical steps, backtrack from unproductive exploration, preserve exact substrings, or remove irrelevant steps.

The authors instantiate this framework in LongSeeker, fine-tuning Qwen3-30B-A3B on 10,000 synthesized trajectories generated by DeepSeek-V3.2. Evaluation covers BrowseComp, BrowseComp-ZH, xbench-2505, and the text-only subset of GAIA. The paper reports 61.5% on BrowseComp and 62.5% on BrowseComp-ZH under its main evaluation configuration. In a controlled comparison using the same training questions, teacher model, base model, and SFT configuration, Context-ReAct-SFT improves BrowseComp from 43.5% to 56.0%, BrowseComp-ZH from 56.0% to 57.3%, and GAIA-text from 74.5% to 77.7%, while leaving xbench unchanged at 78.0%. The authors also report a 43.4% reduction in total token consumption on a 200-question BrowseComp sample, despite more tool calls.

The main contribution is a general agentic paradigm for elastic context orchestration, with five meta-operations forming a fine-grained action space for dynamically managing working context.

Critical strengths
  1. A unified and operationally explicit approach to context management. The five operations provide a clear action space for handling different types of historical information, including exact evidence that should not be paraphrased and unproductive branches that should be abandoned. This is a useful design for long-horizon search agents, where indiscriminate context accumulation can degrade efficiency and reasoning quality.
  2. A controlled comparison provides evidence of a meaningful performance improvement. Under a matched training setup, Context-ReAct-SFT improves BrowseComp by 12.5 percentage points and GAIA-text by 3.2 points relative to ReAct-SFT. The reported reduction in total token consumption also suggests that context management may improve efficiency rather than merely trading a larger context budget for higher accuracy. These results support further investigation of learned context orchestration.
Critical weaknesses
  1. Limited conceptual novelty over existing context-management approaches. The proposed Context-ReAct framework combines five operations—Skip, Compress, Rollback, Snippet, and Delete—into a unified action space for context management. However, the conceptual distinction from existing proactive context-management methods, particularly AgentFold and ARC, is not sufficiently clear.

  2. Insufficiently developed formulation of context management as a learned agent policy. The paper argues that an agent should adapt the fidelity and structure of its working context according to task state, but this principle remains largely intuitive.

Questions for authors
  1. The paper identifies limitations of existing context-management approaches and proposes five fine-grained operations. However, why can these limitations not be addressed by straightforward combinations or extensions of existing methods such as AgentFold and ARC? What is the fundamental conceptual novelty of Context-ReAct beyond a more comprehensive operation set?

  2. What is the principled rationale for selecting these five operations? Are they designed to address distinct and necessary failure modes, and how does the paper establish that this particular set is more appropriate than simpler alternatives?

  3. The paper claims that Context-ReAct is a general agentic paradigm applicable beyond search tasks, but the current evaluation focuses exclusively on search-related benchmarks. Could the authors provide a simple proof-of-concept experiment on a different type of long-horizon agent task to demonstrate whether the proposed context-management framework generalizes beyond search?

  4. Can the authors provide more comprehensive ablations isolating the contributions of individual operations and their combinations? In particular, how much of the improvement comes from specific operations rather than context management in general?

5 comments

Thank you for your assessment of our work and for recognizing the value of our unified context-management framework and controlled experiments. We appreciate your constructive suggestions and address your questions below.

1. [Q1] Conceptual Novelty Beyond Existing Context-Management Methods

We agree that individual context-management operations are not entirely new. Our novelty lies in the unified Context-ReAct paradigm rather than any single operation. Unlike AgentFold's suffix-based folding, Context-ReAct supports editing arbitrary parts of the reasoning history. Compared with ARC's separate context manager, Context-ReAct integrates fine-grained context-management operations directly into the agent's reasoning and action-generation process.

Context-ReAct jointly generates reasoning, multiple meta-operations, and standard tool calls in a single autoregressive pass. Learned through SFT, this unified, composable framework enables the agent itself to decide when, where, and how to manage context.

2. [Q2] Rationale for the Five Meta-Operations

The five operations address distinct context-management needs in long-horizon reasoning:

  • Skip: Preserves context when no modification is needed.
  • Compress: Summarizes resolved history to reduce token consumption.
  • Rollback: Removes unproductive branches and restores earlier context.
  • Snippet: Preserves exact evidence that abstractive summaries might distort.
  • Delete: Selectively removes irrelevant or low-value steps.

Together, these operations balance context efficiency, information fidelity, and reasoning structure.

Our operation-level ablations (Q4) support their complementary contributions. Removing Compress, Rollback, Snippet, and Delete reduces BrowseComp accuracy by 9.2, 3.8, 1.2, and 0.9 pp, respectively. The full configuration achieves the highest accuracy, supporting the value of the combined operation space.

3. [Q3] Generalization Beyond Web Search

To evaluate generalization beyond web search, we additionally evaluate ReAct-SFT and Context-ReAct-SFT on FrontierScience-Research, a benchmark of 60 PhD-level scientific research tasks spanning physics, chemistry, and biology, under matched settings.

MethodFrontierScience-Research Accuracy (%)
ReAct-SFT18.9
Context-ReAct-SFT21.1

Context-ReAct improves accuracy by 2.2 pp, providing evidence that learned context orchestration transfers beyond web search to complex scientific reasoning tasks.

4. [Q4] Contribution of Individual Operations

To evaluate the contribution of each meta-operation, we conduct ablations on the trained LongSeeker-30B-SFT model using BrowseComp. Each operation is disabled individually while keeping the same checkpoint and evaluation settings.

MethodBrowseComp (%)Δ (pp)Total Tokens (K)Token Reduction (%)
ReAct-SFT43.5−12.53,271.3—
Full Context-ReAct-SFT56.0—1,622.850.4%
w/o Compress46.8−9.22,741.316.2%
w/o Rollback52.2−3.81,982.939.4%
w/o Snippet54.8−1.21,874.542.7%
w/o Delete55.1−0.91,902.841.8%

Compress and Rollback have the largest impact on accuracy, while Snippet and Delete provide smaller gains. Full Context-ReAct achieves the highest accuracy with a 50.4% token reduction relative to ReAct-SFT. These results support the complementary contributions of fine-grained context-management operations to both task performance and token efficiency.

Thank you again for your valuable feedback. We hope these additional experiments and clarifications address your concerns, and we would greatly appreciate it if you could increase your rating.

Thanks to the author for a thorough response and further experiments. Although new experiments answer our doubts, I think the paper is flawed in updating key concepts, and a key difference from existing methods is not clear. So, after careful consideration, I decided to keep the original score

Thank you for your follow-up and careful consideration. We would like to further clarify that the key novelty of Context-ReAct lies in treating context orchestration as an integral part of the agent's reasoning and action policy, rather than simply introducing additional context-management operations.

To make this distinction more concrete, we summarize the capabilities of closely related methods below.

✓: Native support; △: Partial or indirect support; ✗: No corresponding native mechanism.

MethodArbitrary-range CompressLossless SnippetSelective DeleteStructural RollbackComposable Meta OpsJoint Meta Ops + Tool Generation
AgentFold✗✗✗△✗✓
ARC△✗△△✗✗
Sculptor✓△△✗△✗
Context-Folding✗✗✗△✗✗
MemAct△✗✓△△✗
Context-ReAct (Ours)✓✓✓✓✓✓

Our novelty lies in the unified Context-ReAct paradigm rather than any single operation. Unlike AgentFold's suffix-based folding or Context-Folding's predefined branches, Context-ReAct supports editing arbitrary parts of the reasoning history. Compared with ARC's separate context manager, Sculptor's fragment-based tools, and MemAct's memory editing actions, our framework integrates lossless evidence extraction, selective deletion, and structural rollback into a composable operation space.

More importantly, Context-ReAct jointly generates reasoning, multiple meta-operations, and standard tool calls in a single autoregressive pass. Learned through SFT, this enables the agent itself to decide when, where, and how to manage context, realizing our central principle of state-dependent context fidelity.

Our controlled ReAct-SFT comparison and additional operation-level ablations further demonstrate the effectiveness of this unified framework. We hope these clarifications address your concerns regarding the conceptual novelty and distinctions from existing methods.

Reviewer_iwjtOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Context-ReAct, an agent paradigm that jointly generates reasoning, context-management operations, and tool actions. It defines five operations—Skip, Compress, Rollback, Snippet, and Delete—so the agent can leave context unchanged, summarize resolved history, preserve exact snippets, remove low-value steps, or roll back an unproductive branch. The authors instantiate this paradigm in LongSeeker, fine-tuning Qwen3-30B-A3B on 10K synthesized search trajectories. The model is evaluated on BrowseComp, BrowseComp-ZH, xbench-2505, and GAIA-text. It reports 61.5 on BrowseComp and 62.5 on BrowseComp-ZH in the main setting (56.0 and 57.3 without discard-all), with 78.0 on xbench-2505 and 77.7 on GAIA-text. A matched ReAct-SFT comparison uses the same base model, teacher, training questions, and SFT configuration, and shows gains of 12.5 points on BrowseComp, 1.3 on BrowseComp-ZH, and 3.2 on GAIA-text. On a 200-question BrowseComp sample, Context-ReAct-SFT uses 43.4% fewer total tokens despite making more tool calls.

Critical strengths
  1. The context-operation space is expressive and easy to understand. The five operations cover several distinct needs: preserving exact evidence, compressing resolved content, removing failed branches, and avoiding unnecessary edits. In particular, Snippet addresses cases where abstractive summaries might distort numbers, names, quotations, or code.
  2. The paper provides a controlled comparison against ReAct-SFT. Using the same base model, teacher, training questions, and SFT configuration is a useful attempt to isolate the effect of the Context-ReAct scaffold and its resulting supervision trajectories. The gains on BrowseComp and GAIA-text are encouraging.
  3. The benchmark coverage is broader than a single search dataset. The evaluation includes English and Chinese BrowseComp as well as xbench-2505 and GAIA-text, giving some evidence that the approach can transfer beyond one exact benchmark format.
  4. The paper measures token use as well as answer quality. The reported 43.4% reduction in total tokens on the 200-question sample is relevant to the motivation. It is also useful that the paper reports the increase in tool calls, rather than presenting token savings without that trade-off.
  5. The method is operationally concrete. Context edits are integrated into the agent loop and applied before the next observation is appended, making the proposal more specific than a general recommendation to “manage memory better.”
Critical weaknesses
  1. The paper does not fully isolate the contribution of individual operations. The matched ReAct-SFT baseline supports the value of Context-ReAct supervision as a whole, but it does not show which operations are necessary for the gains. Ablations removing Rollback, Snippet, Delete, or Compress—and comparisons with a simpler subset of operations—would clarify whether the full action vocabulary is justified.
  2. The headline comparison with other agents is not fully controlled. Table 1 includes systems with different training pipelines (including CPT, SFT, and RL), and some numbers are sourced from official publications or public evaluation platforms. These results provide context but do not establish that context orchestration alone explains the performance gap. The matched ReAct-SFT experiment is more informative and should receive greater emphasis in the conclusions.
  3. The discard-all setting needs careful interpretation. The paper reports 61.5/62.5 on the two BrowseComp benchmarks but also 56.0/57.3 without discard-all. Because discard-all changes how the system handles context, the headline results may not represent the performance of context management alone. The paper should clarify exactly when discard-all is permitted, whether it affects comparable baselines, and how it changes the task-solving procedure.
  4. The evaluation of context efficiency is relatively limited. The token analysis is based on 200 sampled BrowseComp questions. It would be useful to report distributions, variance, and results across the other benchmarks, including context length over time, compression frequency, and the proportion of trajectories where each operation is used. More tool calls alongside fewer tokens also makes end-to-end latency and monetary cost important.
  5. The training-data and synthesis pipeline are central but deserve more scrutiny. LongSeeker is fine-tuned on 10K synthesized trajectories. The quality of these trajectories and the teacher's behavior may substantially influence the results. The paper should better characterize how trajectories are generated, filtered, and validated, and how much of the improvement comes from the context-operation schema versus differences in the generated demonstrations.
  6. The generality of the paradigm remains uncertain. The method is trained as an integrated agent and evaluated primarily on search-oriented tasks. It is not clear whether the same operations and supervision strategy would work for other long-horizon settings such as coding, office workflows, or environments with irreversible actions and persistent constraints.
  7. The paper's novelty relative to concurrent context-management methods could be sharpened. The flexible operations and unified action loop are a coherent design, but the paper would benefit from a more concrete comparison with methods that also learn or expose context edits, especially on the dimensions of operation granularity, training objective, and whether the task agent itself must be fine-tuned.
Questions for authors
  1. Can you provide operation-level ablations (e.g., without Rollback, Snippet, Delete, or Compress) and quantify each operation's effect on task success and token use?
  2. What exactly is the discard-all setting, under what conditions is it used, and are the corresponding baseline numbers evaluated under the same protocol?
  3. Can you report uncertainty estimates and token/latency/cost distributions over multiple runs and across all four benchmarks, rather than only a 200-question BrowseComp sample?
  4. How are the 10K synthesized trajectories generated and filtered? What evidence indicates that the improvement is not mainly due to teacher-generated trajectory quality or other differences in the SFT data?
  5. Can you compare Context-ReAct more directly with recent learned context-management approaches using matched base models, training budgets, and evaluation protocols?
  6. Does the approach retain its gains on tasks outside web search, especially settings involving persistent constraints, multi-step tool use, or irreversible actions?
6 comments

Thank you for your assessment of our work and for recognizing the value of our fine-grained operation space, controlled comparison, and token efficiency. We appreciate your constructive suggestions and address your questions below.

1. [Q1] Contribution of Individual Operations

To evaluate the contribution of each meta-operation, we conduct ablations on the trained LongSeeker-30B-SFT model using BrowseComp. Each operation is disabled individually while keeping the same checkpoint and evaluation settings.

MethodBrowseComp (%)Δ (pp)Total Tokens (K)Token Reduction (%)
ReAct-SFT43.5−12.53,271.3—
Full Context-ReAct-SFT56.0—1,622.850.4%
w/o Compress46.8−9.22,741.316.2%
w/o Rollback52.2−3.81,982.939.4%
w/o Snippet54.8−1.21,874.542.7%
w/o Delete55.1−0.91,902.841.8%

Compress and Rollback have the largest impact on accuracy, while Snippet and Delete provide smaller gains. Full Context-ReAct achieves the highest accuracy with a 50.4% token reduction relative to ReAct-SFT. These results support the complementary contributions of fine-grained context-management operations to both task performance and token efficiency.

2. [Q2] Discard-all Protocol and Fair Comparison

Following MiroThinker, we adopt discard-all which resets the interaction history when a rollout reaches the 200-step limit without producing an answer, allowing up to five rounds.

Table 1 provides a reference for LongSeeker's performance relative to existing systems. For BrowseComp and BrowseComp-ZH, baseline scores follow their respective reported evaluation protocols. Scores marked with * denote ReAct-based results without context management, for which corresponding context-managed results were unavailable.

To ensure a fair comparison, we additionally evaluate ReAct-SFT under the same discard-all protocol as Context-ReAct-SFT, keeping the base model, training data, and SFT settings unchanged.

MethodBenchmarkw/o Discard-allw/ Discard-all
ReAct-SFTBrowseComp43.550.4
BrowseComp-ZH56.060.9
Context-ReAct-SFTBrowseComp56.061.5
BrowseComp-ZH57.362.5

Context-ReAct consistently outperforms ReAct under both settings. With matched discard-all, it retains gains of 11.1 pp on BrowseComp and 1.6 pp on BrowseComp-ZH, demonstrating that its improvements, particularly on BrowseComp, cannot be attributed solely to additional retries.

3. [Q3] Evaluation Uncertainty and Inference Cost

For BrowseComp, LongSeeker achieves 61.5% accuracy, with a 95% run-level confidence interval of [60.4%, 62.4%].

Context-ReAct reduces total token consumption by 50.4% compared with ReAct-SFT (1,622.8K vs. 3,271.3K tokens per trajectory), despite more tool calls (129.48 vs. 67.91). These statistics are based on all 400 BrowseComp questions without discard-all. Since BrowseComp involves longer-horizon searches than the other three benchmarks, it provides a more representative setting for evaluating the benefits of context management.

Regarding cost and latency, model inference dominates the computational and monetary cost, while Search/Visit API calls are relatively inexpensive. Wall-clock time is affected by API latency fluctuations and concurrent execution, making direct comparisons unreliable. We therefore use total token consumption as the primary measure of inference efficiency.

4. [Q4] Training Data Synthesis and Quality

As described in Sec. 3.3, we collect 10K questions from OpenSeeker (9K English and 1K Chinese) and use DeepSeek-V3.2 to generate Context-ReAct trajectories, retaining trajectories with valid output formats.

63% of the 10K teacher trajectories reach a correct final answer. We also evaluated rejection sampling by retaining only correct trajectories, but BrowseComp accuracy dropped from 56.0% to 53.5%. One possible explanation is that trajectories with incorrect final answers still provide valuable supervision for long-horizon exploration and context-management strategies.

Importantly, our controlled comparison in Table 2 uses the same teacher, base model, training questions, and SFT configuration for ReAct and Context-ReAct, isolating the effect of the agent scaffold and its resulting supervision trajectories.

5. [Q5] Comparison with Related Context-Management Methods

We agree that direct comparisons with recent learned context-management methods are important. However, a fully matched comparison is computationally infeasible, as existing methods differ in their base models, training pipelines, and context-management mechanisms, requiring separate reimplementation, method-specific trajectory synthesis, and retraining under matched budgets.

Regarding W7 on novelty, Context-ReAct differs in its unified reasoning paradigm and fine-grained operation space. We clarify these distinctions below. We clarify these distinctions below.

✓: Native support; △: Partial or indirect support; ✗: No corresponding native mechanism.

MethodArbitrary-range CompressLossless SnippetSelective DeleteStructural RollbackComposable Meta OpsJoint Meta Ops + Tool Generation
AgentFold✗✗✗△✗✓
ARC△✗△△✗✗
Sculptor✓△△✗△✗
Context-Folding✗✗✗△✗✗
MemAct△✗✓△△✗
Context-ReAct (Ours)✓✓✓✓✓✓

Our novelty lies in the unified Context-ReAct paradigm rather than any single operation. Unlike AgentFold's suffix-based folding or Context-Folding's predefined branches, Context-ReAct supports editing arbitrary parts of the reasoning history. Compared with ARC's separate context manager, Sculptor's fragment-based tools, and MemAct's memory editing actions, our framework integrates lossless evidence extraction, selective deletion, and structural rollback into a composable operation space.

More importantly, Context-ReAct jointly generates reasoning, multiple meta-operations, and standard tool calls in a single autoregressive pass. These capabilities are learned through SFT, enabling the task agent itself to decide when, where, and how to manage context.

Table 2 provides a controlled comparison with ReAct-SFT under matched training settings, while our additional operation-level ablations examine the contributions of individual operations. These results provide empirical support for Context-ReAct, whereas the capability comparison highlights its methodological distinctions from prior work.

6. [Q6] Generalization Beyond Web Search

To evaluate generalization beyond web search, we additionally evaluate ReAct-SFT and Context-ReAct-SFT on FrontierScience-Research, a benchmark of 60 PhD-level scientific research tasks spanning physics, chemistry, and biology, under matched settings.

MethodFrontierScience-Research Accuracy (%)
ReAct-SFT18.9
Context-ReAct-SFT21.1

Context-ReAct improves accuracy by 2.2 pp, providing evidence that learned context orchestration transfers beyond web search to complex scientific reasoning tasks.

Thank you again for your valuable feedback. We hope these additional experiments and clarifications address your concerns, and we would greatly appreciate it if you could reconsider your rating.

Comments on the paper

1 comment

Dear Reviewers and Area Chair,

We sincerely thank all reviewers for their careful reading and constructive feedback. We have conducted additional experiments and provided detailed responses addressing the main concerns regarding operation-level contributions, evaluation fairness, inference efficiency, methodological novelty, training data quality, and generalization. We summarize our key findings below.


1. Contribution of Individual Meta-Operations

We conducted operation-level ablations under matched evaluation settings. The results demonstrate the complementary contributions of the five meta-operations, with Compress and Rollback playing particularly important roles in improving task performance and context efficiency.


2. Fair Comparison and Inference Efficiency

We clarified the discard-all protocol and evaluated ReAct-SFT and Context-ReAct-SFT under matched inference settings. Context-ReAct consistently maintains its advantage, confirming that its improvements are not solely attributable to additional retries.

We also extended our token analysis, demonstrating substantial reductions in token consumption, while separately quantifying the additional inference cost introduced by discard-all.


3. Methodological Novelty and Training Data Quality

We clarified Context-ReAct's distinctions from recent context-management methods, emphasizing its unified, composable paradigm that jointly generates reasoning, meta-operations, and tool calls.

We further analyzed teacher-generated trajectories and conducted correctness-based rejection sampling experiments, providing additional evidence on the effectiveness of our training approach.


4. Generalization Beyond Search

We conducted additional experiments on complex scientific research tasks, demonstrating that Context-ReAct's benefits extend beyond the original search benchmarks.


We hope these additional experiments and clarifications address the reviewers' main concerns and strengthen the empirical support for Context-ReAct. We would appreciate it if the Area Chair could encourage further reviewer engagement while there is still time for discussion. We welcome any remaining questions and would be happy to provide further clarification. We also sincerely hope the reviewers will reconsider their assessments and increase their ratings accordingly.

Thank you again for your time and consideration.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.