acceptodds
Under review as a conference paper at ICLR 2027

Adaptive Organization-Centric Collaboration for LLM-Based Multi-Agent Systems

Abstract

Large language model (LLM)-based multi-agent systems (MAS) excel at complex tasks, and their effectiveness depends critically on how agents are organized to collaborate. However, most existing methods design collaboration only at the instance level, generating for each query a concrete topology that fixes which agents participate and how they communicate before execution. They leave unmodeled the organizational level of collaboration, which comprises the groups through which agents collaborate, how these groups form and change as the task unfolds, and how context passes between them. As a result, these methods cannot adapt collaboration during execution, and the organizational knowledge they learn stays tied to whole-task topologies. Inspired by the Agent-Group-Role (AGR) organizational model, which separates the organizational structure from the concrete organizations that instantiate it, we lift collaboration design to the organizational level and propose AGR-AR, an organization-centric framework. AGR-AR learns how to organize agents and instantiates groups on demand during execution. At each step, a lightweight autoregressive topology generator constructs the next group conditioned on the task and execution feedback, selecting its members, including overlapping agents that carry context from earlier groups, and determining how they communicate. The organization thus adapts during execution, and a single generator serves tasks across domains. Experiments on six benchmarks show that AGR-AR achieves higher average accuracy than strong baselines while consuming substantially fewer tokens and remaining more robust to adversarial agents; without retraining, it also incorporates unseen agents and transfers to unseen tasks. Code is available at https://anonymous.4open.science/r/AGR-AR-ECF6/.

Primary area:
foundation or frontier models, including LLMs
Keywords:
large language models, multi-agent systems, LLM agents, multi-agent collaboration, graph generation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_1agsOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper argues that most existing topology design methods for LLM-based multi-agent systems typically determine a complete collaboration topology for a task before execution and keep it largely fixed throughout the subsequent process. The authors argue that, rather than directly learning “what complete topology should correspond to a given task,” it is more desirable to learn a more general principle of “how agents should be organized under different task states.” To this end, they train a lightweight topology generation model that, based on the current task, execution feedback, and historical information, dynamically generates the next group of agents and their communication topology on demand during task execution, allowing the collaboration structure to adapt as the task progresses.

Critical strengths
  1. The idea of dynamically adapting the collaboration topology is very interesting.
  2. The experiments are comprehensive and cover multiple benchmarks and ablation studies.
  3. The story is insightful and made me rethink the design of multi-agent collaboration topologies.
  4. The method is lightweight while still achieving strong performance.
Critical weaknesses
  1. The necessity of using such a small model for topology selection is unclear, especially since GPT-4o already understands the current execution state.
  2. The performance gain is small, and it is unclear whether it is robust or truly caused by dynamic topology adaptation.
  3. The ablation results suggest that group decomposition and cross-group overlap contribute more than dynamic adaptation, which may not fully support the paper’s main story.
  4. Dynamically changing agent roles across groups may introduce context or role confusion, especially for weaker models.
Questions for authors

Thank you for your paper. You propose a very interesting idea, but I have several questions regarding the paper.

  1. Is it sufficient to use such a small model to determine the topology? Small models are generally more likely to make mistakes than large models. Letting such a small model take the role of topology selection makes me wonder whether this could introduce problems into the results. Is it possible that this small model is actually not that important, since GPT-4o already understands what is currently happening?

  2. The performance improvement is quite small. It is unclear whether this is due to variance, and it is also unclear whether this small gain really comes from the dynamic topology.

  3. According to the ablation study, removing dynamic adaptation causes an average drop of only 1.76 points, while removing cross-group overlap causes a drop of around 4.5 points. Then how can we tell whether the gain really comes from dynamic adaptation? Is it possible that the improvement mainly comes from splitting the collaboration process into multiple groups? Therefore, I think there may be a gap between the paper's story and its experimental results.

  4. Could dynamically changing the topology increase context confusion for the LLM agents? Is it possible that the same model plays opposite roles at different time steps, for example, acting as a proposer at one step and an opponent at the next? For less capable models, could this framework therefore introduce context confusion? If the context needs to be reset between groups, then I think this method looks more like splitting a large task into several smaller tasks and then pre-determining a collaboration topology for each small task. This seems somewhat different from the story proposed in the paper, since the paper argues that the communication topology should change dynamically rather than being determined in advance.

Reviewer_rP5TOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

AGR-AR organizes LLM agents into a sequence of groups generated during execution. A lightweight autoregressive model selects agents and communication edges using task context and feedback, while overlapping agents carry information between groups. The paper reports improvements in accuracy and token efficiency across six benchmarks, with additional transfer experiments.

Critical strengths
  1. The motivation and design are well aligned: generating groups incrementally allows collaboration to respond to intermediate results, while overlapping agents preserve continuity.
  2. The reported accuracy–cost tradeoff across reasoning and coding tasks is promising. The released implementation also makes the proposed execution framework easier to examine.
Critical weaknesses
  1. The generator’s contribution is not sufficiently isolated. The system combines learned group generation with LLM-based mission planning and dataset-specific role restrictions. These restrictions also influence the activation patterns used to support organizational generalization. More controlled comparisons are needed to establish the benefits of the proposed generator.
  2. Implementation discrepancies limit confidence in the results. Static code inspection shows that the overlap loss provides no gradient and that edge features differ between training and inference. HumanEval test invocation and the procedure for adding new agents also require clarification. It remains unclear whether the released code matches the experimental version. The findings from inspecting the released code have reduced my confidence in the paper’s claims and reported results.
Questions for authors
  1. Does the released implementation match the version used in the experiments? How were the overlap loss and training–inference feature mismatch handled? Does the processed HumanEval test field explicitly invoke check(entry_point)? A representative processed record would resolve this concern.
  2. How does AGR-AR compare with fixed group templates using the same LLM planner, role pool, and inference budget? Were the dataset-specific role restrictions also applied to baselines?
  3. How was the fixed-size overlap projection adapted when adding the four new agents for GAIA without retraining?
Reviewer_k8t6OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses the limitation that existing large language model (LLM)-based multi-agent systems rely excessively on static, instance-level topology design, and proposes AGR-AR, a dynamic collaboration framework based on an organizational-level perspective. Inspired by Agent-Group-Role (AGR) theory, the framework no longer predefines a complete global communication graph for each query. Instead, during execution, it uses a lightweight autoregressive graph generator to instantiate local agent groups on demand, conditioned on the task context and real-time execution feedback. To ensure contextual continuity, the system passes necessary information by reusing agents across adjacent groups (overlapping agents). The generative model is trained progressively in three stages—structural prior learning, task-topology alignment, and efficiency-guided refinement—to achieve high accuracy and low communication cost across multiple benchmarks.

Critical strengths
  1. The paper's paradigm shift from instance-level static topologies to organizational-level dynamic group instantiation is logically coherent and fundamentally alleviates the pain point that predefined topologies cannot adapt during execution.
  2. The efficiency analysis (token consumption) in the experiments is well-supported. The results on GSM8K clearly demonstrate that on-demand group instantiation can substantially reduce redundant communication, achieving the highest accuracy while consuming very few tokens.
Critical weaknesses
  1. The authors cite the latest 2026 test-time dynamic adaptation methods (e.g., MAS-on-the-Fly) and the single-model cross-domain method (OFA-MAS) in Sections 2.1 and 2.2, yet exclude all of them from the main experiments (Table 1), retaining only older or static baselines. This appears to deliberately avoid direct competitors.
  2. The paper claims that the model autonomously learns "organizational knowledge," but its Stage 2 and Stage 3 training data are entirely generated and validated by GPT-4o. This essentially distills GPT-4o's strong prior knowledge into a 2.9M model. The "extremely low token consumption" advertised in Figure 3a also deliberately conceals the enormous API training cost of Stage 2. Comparing against baselines that are not supervised by a teacher model of comparable scale seriously undermines the rigor of the experiments.
  3. Sensitivity analysis of core hyperparameters is severely missing. The cost weight β\beta declared in Eq. (4) and the loss weights α=0.2\alpha=0.2 and λ=0.1\lambda=0.1 in Eq. (12) have no sensitivity validation whatsoever; Table 2 further avoids the key "random grouping" control experiment. In addition, Section 5.5 forcibly attributes the improvement in adversarial robustness to "group boundary isolation" without ruling out the possibility that the model simply avoids selecting malicious nodes during generation via Eq. (9). This attribution lacks strict variable control.
  4. Stripping away the grand AGR theoretical packaging, the underlying graph generation algorithm is outdated. The overlap vector in Eq. (2) is merely a coarse 0/1 binary marker that erases cross-group historical identity information; the entire autoregressive generator is essentially just the 2018 GraphRNN framework combined with basic dot products and MLPs, lacking substantive low-level innovation for multi-agent graph structures.
Questions for authors
  1. The authors cite strong competing methods from the same year, 2026, in Section 2 (e.g., MAS-on-the-Fly, AgentConductor, and OFA-MAS). Why are these state-of-the-art baselines with test-time dynamic adaptation or cross-domain capabilities not compared in the main experiments in Table 1?
  2. Why is there no sensitivity analysis supporting the cost weight β\beta defined in Eq. (4), or the core loss weights α=0.2\alpha=0.2 and λ=0.1\lambda=0.1 in Eq. (12)?
  3. In the adversarial robustness experiments in Section 5.5, the authors attribute performance retention to “group boundary isolation.” How can it be demonstrated that this is not simply because the autoregressive generator (Eq. 9) directly avoids malicious nodes during sampling?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.