- 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
-
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.
-
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
-
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.
-
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
-
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.
-
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?
-
Do you have any evidence from the trained model on the contribution of individual operations, for example from disabling Snippet or Rollback at inference?
-
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?
-
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
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.
| Method | Arbitrary-range Compress | Lossless Snippet | Selective Delete | Structural Rollback | Composable Meta Ops | Joint 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
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.
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.
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.