acceptodds
Under review as a conference paper at ICLR 2027

GateTree: Gated Context Sharing for Adaptive Multi-Agent Collaboration

Abstract

Parallel multi-agent systems share intermediate findings to coordinate different solution paths. As these paths evolve, however, shared information can become outdated or inapplicable, propagating unsupported assumptions and causing redundant work. Effective collaboration requires selecting information that remains relevant to each branch and adapting its plan accordingly. We introduce GateTree, a multi-agent framework that couples shared-context management with adaptive planning. GateTree maintains two concurrent branches with complementary conservative and exploratory strategies. A GateAgent checks the provenance and scope of shared records and selects context for each branch. An adaptive planner then uses this context and execution feedback to revise branch objectives and verification targets, preserving useful progress while addressing unresolved requirements. Across SWE-bench Pro, text-only HLE, and GAIA2, GateTree achieves the highest average Pass@2 under both backbone models while reducing cost by an average of 30.17% relative to each setting's strongest Pass@2 baseline. Replacing active context management with passive sharing reduces Pass@2 by 11.50 points in the component ablation. Further analysis shows that selective delivery is most beneficial for long contexts with sparse relevant evidence, whereas full-context delivery can be more effective when relevant evidence is dense.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Multi-agent systems, shared-context management, adaptive planning, selective information sharing, large language models

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

This paper presents GateTree, a multi-agent collaboration framework that integrates selective context sharing with adaptive planning. GateTree maintains two concurrent branches with complementary conservative and exploratory strategies. A GateAgent manages shared information through provenance tracking, applicability filtering, and recipient-specific context selection, while an adaptive Planner updates branch objectives and verification requirements based on shared evidence and execution feedback.

The authors evaluate GateTree on SWE-bench Pro, text-only HLE, and GAIA2 using two backbone models. GateTree achieves the highest benchmark-averaged Pass@2 across the compared methods, with an average inference cost reduction of 30.17% relative to the strongest Pass@2 baseline in each setting. The paper also presents extensive component ablations, context-selection studies, and analyses of agent coordination.

Critical strengths
  1. The research question is critical and practical. Managing shared information across concurrent agents is an important challenge, particularly when evidence becomes stale, contradictory, or irrelevant as different branches evolve. The paper identifies a meaningful problem in multi-agent collaboration.

  2. The experiments are comprehensive. The experiments cover three substantially different benchmarks and two backbone models. GateTree consistently achieves competitive or superior task-level coverage while maintaining relatively low inference costs.

  3. The controlled experiments varying context length and relevant-evidence density provide useful insights into when selective context delivery is beneficial.

Critical weaknesses
  1. The overall novelty is limited, and the differentiation from prior work is insufficient. The proposed framework largely combines established techniques, including structured shared memory, provenance-based filtering, recipient-aware context selection, heterogeneous agent roles, and feedback-driven replanning. In particular, RCR-Router [1] already investigates role- and task-stage-aware context routing over structured shared memory, with iterative memory updates and token-budget constraints. The paper also builds on ideas closely related to DeLM's feedback-driven coordination and AgentFugue's selective sharing of peer information. Although GateTree introduces more explicit applicability checks and plan-conditioned evidence delivery, it remains unclear whether these extensions constitute a substantial methodological advance.

  2. The paper compares against several established agent frameworks, including DeLM and AgentFugue. However, it lacks a sufficiently strong, lightweight baseline that directly tests whether the proposed architecture is necessary. For instance, a simpler framework combining provenance-aware shared memory, plan-conditioned embedding retrieval, and the same adaptive Planner. Moreovew, although Section 5.2 compares GateAgent against Recent, Priority, and All Context selection, Recent and fixed-priority ranking are relatively weak retrieval strategies. These comparisons do not sufficiently establish the advantage of the proposed LLM-based selection mechanism over a competent semantic-retrieval baseline. Similarly, AOrchestra is discussed in related work but not experimentally compared, despite its relevance to adaptive agent orchestration.

  3. The paper provides meaningful ablation evidence. However, the results do not establish whether the improvements arise from a genuinely distinctive coordination mechanism or simply from combining better information retrieval with standard feedback-driven replanning. For instance, conditioning retrieval on an updated plan naturally increases the relevance of delivered evidence. It remains unclear whether a standard plan-aware retrieval mechanism would recover most of the observed improvement. Furthermore, removing adaptive planning causes a substantial 16.5-point reduction in Pass@2, suggesting that a considerable portion of the gains may come from feedback-driven planning, which is already well established.

  4. The paper emphasizes Pass@2 as its primary performance metric. However, improvements in task-level coverage do not consistently translate into higher average success rates across individual submissions. A more detailed analysis separating individual success, solution diversity, and coordination benefits would help clarify the source of the improvements.

[1] RCR-Router: Efficient Role-Aware Context Routing for Multi-Agent LLM Systems with Structured Memory, https://arxiv.org/abs/2508.04903

Questions for authors
  1. To what extent do improvements arise from complementary branch behavior rather than improved individual execution quality? Could the authors provide comparisons against independent conservative/exploratory attempts under matched token budgets and verification procedures?
3 comments

Dear Reviewer,

Thank you again for your constructive feedback. We have provided detailed responses to your concerns, including additional experiments and analyses. As the discussion period approaches its deadline, we would appreciate it if you could review our responses. We would be happy to address any remaining questions or concerns.

Thank you for your time and consideration.

Thanks for your response. Some of my concerns are addressed and I would increase the score to 2.

Thank you for your feedback and for reconsidering your score. We sincerely appreciate it.

If you have any remaining questions or concerns, we would be happy to provide further clarification. We hope our responses and additional evidence may support further reconsideration of the score.

Thank you again for your time and constructive feedback.

Reviewer_Q6MEOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper focuses on an important plan sync problem in multi-agent collaborations. It proposes GateTree, which couples recipient-specific evidence management with adaptive planning driven by feedback. It includes evidence filtering in terms of provenance and revision, record lifecycle management, and context selection. The paper conducted well-designed experiments to verify not only the main performance of the proposed method, but also the mechanisms and boundaries.

Critical strengths
  1. The focused problem is very important, and the method design from the recipient's perspective is well-motivated and rational. Overall, the evidence governance and adaptive replanning framework are well-designed.
  2. The experiments are very comprehensive: various benchmarks, LLM backbones, ablations, and different settings for method boundaries.
Critical weaknesses
  1. The analysis of the experiments could be deeper and more integrated. From Table 1, the proposed method performs well but not in a no-doubt SOTA way. There might be some features (e.g., difficulty, context-length requirements, task type, etc.) of the benchmarks or tasks that affect the performance. Since the paper has some comprehensive analysis in the following section, it would be good to explain Table 1 in details.
Questions for authors
  1. Could the authors further clarify how the updated branch plan changes the evidence selected for subsequent Worker execution, compared with selection based only on the original task and recipient information?
  2. Are the ablation results from a single run?
5 comments

Dear Reviewer,

Thank you again for your constructive feedback. We have provided detailed responses to your concerns, including additional experiments and analyses. As the discussion period approaches its deadline, we would appreciate it if you could review our responses. We would be happy to address any remaining questions or concerns.

Thank you for your time and consideration.

Thanks for your response. Most of my concerns are addressed. In Q3, you mentioned that repeated ablations are underway, so I would appreciate seeing the results when available.

AuthorsOpenReview ↗

Follow-up on Q3: Ablation results across three independent runs

Thank you for following up. We have completed the two additional ablation runs mentioned in our earlier response. Each configuration was evaluated three times on the same 200 SWE-bench Pro tasks using the same DeepSeek model. Run 1 is the evaluation originally reported in the manuscript. We report the individual runs and the mean ± sample standard deviation below:

ConfigurationRun 1 Avg. / Pass@2 (%)Run 2 Avg. / Pass@2 (%)Run 3 Avg. / Pass@2 (%)Avg. mean ± SD (%)Pass@2 mean ± SD (%)
Full GateTree44.25 / 54.5046.25 / 58.0043.00 / 53.0044.50 ± 1.6455.17 ± 2.57
w/o Applicability Filter40.75 / 49.5042.75 / 52.5039.50 / 47.5041.00 ± 1.6449.83 ± 2.52
w/o Plan-conditioned Selection42.00 / 51.0040.75 / 49.0043.75 / 53.5042.17 ± 1.5151.17 ± 2.25
w/o Both39.50 / 47.0037.75 / 44.5040.75 / 48.5039.33 ± 1.5146.67 ± 2.02
Passive Shared Context38.50 / 43.0037.25 / 41.5040.25 / 46.0038.67 ± 1.5143.50 ± 2.29
No Shared Context34.50 / 40.5036.25 / 43.0033.75 / 39.5034.83 ± 1.2841.00 ± 1.80
Homogeneous-238.50 / 47.0037.00 / 44.5040.25 / 49.0038.58 ± 1.6346.83 ± 2.25
Static Planner33.00 / 38.0030.25 / 35.5033.75 / 41.5032.33 ± 1.8438.33 ± 3.01
Passive + Static30.50 / 35.0028.75 / 33.0032.25 / 38.0030.50 ± 1.7535.33 ± 2.52

Full GateTree achieved the highest three-run mean on both metrics. Removing applicability filtering or plan-conditioned selection reduced mean Pass@2 by 5.33 and 4.00 percentage points, respectively. Removing both reduced it by 8.50 points.

We will include the three-run results in the revised manuscript. If you have any further questions, we would be happy to address them. Thank you again for suggesting this check.

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

This paper presents GateTree, a multi-agent framework that tightly couples shared context governance with adaptive planning across two persistent branches. GateAgent manages record lifecycle, source tracking, and scoping to selectively route receiver-specific evidence, while an adaptive planner updates goals and verification targets based on local execution feedback and peer-generated evidence. Experiments spanning SWE-bench Pro, HLE (text-only), and GAIA2 with GLM-5.2 and DeepSeek-V4-Pro-0813 yield the best average Pass@2 across all three benchmarks, at a fraction of the usual cost. Ablation and controlled studies disentangle the effects of selective context delivery and its interaction with re-planning.

Critical strengths
  1. The paper addresses a practical problem in multi-agent collaboration: shared information may become outdated or no longer apply to another branch. A useful aspect of GateTree is that it connects evidence selection with plan revision. The selected evidence helps the planner decide what to do next, while the updated plan guides further context selection. Checking the source, scope, and revision of shared records makes this process more careful than simply sharing recent messages. The conservative and exploratory branches are also easy to understand, and the ablation results support the value of using different strategies.

  2. The experiments cover three benchmarks and two backbone models, reporting both success rates and inference costs. The ablations support the benefits of combining context management with adaptive planning. The context study also shows when selective sharing helps and when full-context delivery may be better.

Critical weaknesses
  1. Pass@2 does not reveal how sharing affects each branch. On GAIA2 with DeepSeek, GateTree exceeds DeLM in Pass@2 (51.63% versus 47.21%) but has lower Avg. (40.60% versus 43.21%). Higher coverage does not imply better average submission quality. Maybe please report each role’s success and compare otherwise identical configurations with sharing enabled and disabled.
  2. The major concern is the failure-mode taxonomy lacks independent validation. Categories derived from 100 tasks per framework are applied to the same 200-task pool. They closely match GateTree’s design goals and serve both as motivation and evaluation. Despite author checking, no framework-blinded labeling or annotator agreement is reported. The categories should be validated on disjoint trajectories with independent annotation.
  3. There is no direct reliability evaluation on GateAgent’s decisions. Table 3 reports downstream benefits, but not invalid deliveries, evidence omissions, or incorrect lifecycle updates. The context study tests cross-task distraction rather than same-task stale evidence. This gap matters for HLE, which lacks executable feedback. A record-level audit covering same-task revisions and HLE disagreements would better establish reliability.
Questions for authors

Please address the weaknesses.

3 comments

Dear Reviewer,

Thank you again for your constructive feedback. We have provided detailed responses to your concerns, including additional experiments and analyses. As the discussion period approaches its deadline, we would appreciate it if you could review our responses. We would be happy to address any remaining questions or concerns.

Thank you for your time and consideration.

Appreciate your responses. I think they fully address my concerns and have raised the score. Please add them into the final version.

Thank you for your positive feedback and for raising the score. We sincerely appreciate your time and suggestions. We will incorporate the clarifications into the final version.

Comments on the paper

1 comment
Area_Chair_SHNNOpenReview ↗

Reminder: Reviewer Engagement During the Discussion Period

Dear Reviewers,

As the discussion period is drawing to a close, could you please review the authors' rebuttal and provide your feedback at your earliest convenience?

Active engagement in discussions with the authors is an essential part of the peer-review process and plays an important role in ensuring the success of the conference. Your participation would be greatly appreciated.

Thank you for your time and contributions.

Best regards,

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.