acceptodds
Under review as a conference paper at ICLR 2027

The Shape of Cooperation Across Tasks

Abstract

When should a team of agents communicate through a hub, through groups or directly? We characterize these choices by two fitted summaries of evidence demand, the traffic outside the best hub and the smallest modular detour. Under a faithful transport model that preserves evidence delivery, this signature and the ratio of channel overhead to traffic cost determine the cheapest pattern within the declared family. Identical hub concentration can require different patterns, so concentration alone is insufficient. We derive exact decision boundaries and a confidence bound that charges for learning before authorizing a change. Controlled transport experiments show that matching connectivity to persistent demand can lower total modeled cost after paying to learn it. A direct ablation shows that access to modular patterns improves future cost for concentrated demand but adds no benefit for diffuse demand. More observations improve prediction but can reduce net savings, while replacing future demand reverses the benefit. The resulting characterization explains when fitting a richer communication pattern is useful and when its learning expense or changing demand defeats the gain.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Multiagent systems, autonomous organization, collaboration topology, adaptive coordination, collective intelligence

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 4 reviews, above 9% of the others.
Reviewer_Hr3yOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper models a fixed team forwarding unchanged evidence packets over a directed network. Enabled channels incur session costs; packet hops incur traffic costs. For six workers, it compares complete direct communication, six bidirectional stars, and 90 modular graphs, each with two groups of three and one gateway per group. Two fitted quantities summarize these choices. Hub residual demand measures traffic not incident to the best hub. Modular detour gives the least added forwarding cost after choosing partitions and gateways. Their cost contrasts with direct communication identify the cheapest family. Within a fixed partition, gateways can be selected independently. Statistical bounds cover demand estimates, distribution change, and observation or switching costs.

Critical strengths

I found the problem formulation interesting because it connects how a team is organized to the cost of sharing information. Costs, preservation of delivery, candidate graphs, fitted summaries, and tie rules are explicit. The gateway identity simplifies implementation. Equal-hub-concentration examples show why concentration alone cannot distinguish local from diffuse exchange.

I particularly like the care taken to separate learning from future evaluation. Methods share the same future packets, which are separate from observed requests; development and evaluation seeds also differ. Learning costs are charged even when the graph stays unchanged. The permutation control separates finding the right worker connections from choosing a graph family.

I appreciate the authors' willingness to show where the approach does not help. They retain the null modular result at high channel price, losses under demand replacement, costs of rejected changes, and the stronger selective fixed planning control. They identify the modular ablation and sharper confidence analysis as post hoc and document reasoning as planned work.

Critical weaknesses

My main reservation is what the signature adds beyond the optimized costs of the chosen graph families. The two demand summaries are optimized forwarding penalties of predefined alternatives. Each family has a fixed edge count, so their ability to identify the cheapest family follows from comparing three affine cost expressions.

I can see the modeled savings from matching connectivity to persistent demand after paying learning expenses. What I am not yet convinced of is that the favorable prices describe a real system. The gains depend on assigned channel, collection, selection, and switching prices. Selector CPU time is measured, but neither this nor request acquisition is converted into the units used for session and hop costs.

Questions for authors

What insight or predictive advantage remains beyond fixed-family cost minimization when group counts and sizes can vary? To judge the importance of this restriction, I would also need to know how much savings the two-triad family leaves unrealized.

Could you identify a runtime with a measured per-enabled-channel cost high enough to make the selector useful? Please show whether it still beats selective direct communication after measured acquisition and selection costs.

Please explain what information justifies a deployed policy's drift bound and observation budget. I would then like to see whether a rule fixed before evaluation beats never observing on fresh trials, counting all rejected attempts.

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

The goal of the pair is to understand how communication should be handled between agents that are trying to carry out some tasks in collaboration. How information is passed from one agent to another depends on the task at hand, the type of information, and the connection between the agents. The optimal information passing patterns might be very different based on these factors. The goal of the paper is to explore these patterns. The paper develops a theoretical framework for formalizing and exploring these information patterns. Unfortunately, the writing of the paper is extremely difficult to follow. Basic terms are not clearly defined and links between terms are unclear. The paper gives an example of a software engineering task, but then the rest of the intro and the theoretical frameworks proceeds at a very abstract level which is difficult to comprehend. I have tried to provide action points that the paper might implement to improve the exposition.

Critical strengths

Understanding optimal decision patterns between agents is a timely and interesting problem.

Critical weaknesses
  1. The paper is rather difficult to understand. Going over the intro, the reader gets a rough sense of what the paper is trying to do, but the sentence seem disconnected and quite convoluted. For instance, in line 37, the paper says “forwarding through that worker adds expense when most evidence is needed elsewhere”. It is unclear what “expense” is referring to. What does “most evidence” mean? Do it mean that at least some of the evidence was needed here? The “not merely call the task coding” clause does not seem to follow the sentence that came before it. Terms like “saved connections” and “repay the additional forwarding” in the same paragraph are very difficult to understand. After reading two paragraphs of the intro, the reader still does not know precisely what the underlying problem is. Only some hazy details are understandable at this point.
  2. The figure in the introduction is difficult to parse. The problem setup is unclear. What exactly is the task that we are trying to achieve with various agents? The text is disconnected from the figure. Line 105 mentions the figure but provides not substantive details to help understand it. Similarly, definitions of new terms like “task signature” include vague terms like “fitted hub” and “modular pattern” which were never clearly defined.
  3. The paper makes a connection to Rudin (2019) in the intro and then in the related work. To this reviewer, who considers themselves fairly knowledgeable in the interpretability literature, the connection is unclear. The paper should describe how the message passing between agents relates to interpretability.
Questions for authors

What is the “hidden evaluator label” in line 157?

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

The paper characterizes when agents should communicate via hubs, groups, or some other organization based on evidence from demand features. It further explores if acting on this demand improves efficiency or not. The problem is well founded and prevalent in multi-agent systems. The paper also provides theoretical analysis of how and when certain communication formations are beneficial. The paper does controlled ablation tests on a synthetic environment but omits evaluation in the HotPotQA evaluation framework

Critical strengths
  • The paper is generally well motivated and addresses a common problem in multi-agent systems. I do like the idea of using task evidence demand to motivate organization of communication patterns.
  • I also found the theory to be sound especially with the authors arguments on why hub concentration alone is insufficient to choose an optimal topology, as identical hub concentrations can hide very different modular grouping opportunities.
Critical weaknesses

Content: The paper does not evaluate the method in a common benchmark, nor do they include baselines which makes it hard to see any promise/benefit of the paper's idea compared to other existing methods. There is also a mismatch between the general formulation and the actual experiments which requires clarification. (Please see questions below for more details)

Clarity: I feel like the flow of the paper can be tightened a bit; it was a little hard to follow in the introduction as it felt some of the concepts were introduced with no context (for example Figure 1 is referenced in the intro itself, but doesn't really specify what recorded planning is or what the deal is with selecting a specific "prospective seed"?). There are also many sentences that have a very choppy structure: introduce a point and then immediately argue some other point in the same sentence. Generally the entire paper could do with some improvements in terms of clarity with providing context to new concepts being introduced/new mathematical terms.

Questions for authors

Questions:

  1. How sensitive are the cost boundaries to standard failures in channels and/or nodes? If packets are dropped with some probability, does the advantage of modular/hub topologies collapse compared to direct delivery?
  2. Is there any analysis on if and when the agents deviate from the proposed graph topology (for example, since message sending is optional, what happens if agents don't send a message for some reason or is that not possible in the given settings).
  3. I'm a little confused on the cost structure. As I understand it, it seems like it is for how well a communication structure accomplishes the task minus the cost of reorganizing if required. If so, if the new graph has some communication edge, but the agents don't use that to communicate, could that not potentially affect how well they can accomplish the task? Further, if this was observed, how would the method attribute the issue to the communication pattern versus just agents not utilizing the communication pattern well. Essentially, can the authors clarify how they differentiate between a bad topology vs agents not utilizing a good topology well?
  4. Is session overhead defined as a constant in section 4? If so, could the authors clarify why it can be a constant if the channel is just enabled but no messages pass through it?
  5. General formulation vs actual experiments. If I have understood this correctly, the general formulation is provided for settings with a task score Y, but the setting from the experiments does not use Y anywhere. Further, section 5.4 also specifies that the proposed evaluation for HotPotQA was never executed but all details are still specified. I'm not really clear on why that is the case.
  6. It seems that the executed experiments only compare against internal controls/ablations instead of any of the topology-generation baselines discussed in the related work (e.g., G-Designer, GPTSwarm). Could the authors clarify why no external baselines were included?
Reviewer_zhuvOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper asks how a fixed team of agents should be wired to share information. Its evidence comes from simulations with randomly generated communication patterns and a robot path-planning study using a classic algorithm. No machine learning, neural networks, or language model experiments are used in the paper.

Critical strengths

S1. Generally, I think doing agent simulations to understand social interactions between language models is a promising avenue. I find these social dynamics interesting, however, the paper have fundamental methodological problems.

Critical weaknesses

W1. The major weakness, which I believe is sufficient for the rejection of the paper is the following: As far as I can tell, the paper does not include any experiment where neural networks are involved. Nothing is trained, and the agents that the paper refers to is not language model agents. HotpotQA question-answering study is described as a plan and it was never run. So this is a simulation paper, not a machine learning paper. this is why Ibelieve ICLR is not the right venue for the paper. I will write my review around this theme.

W2. I found it difficult to understand what the paper is actually doing. It took me some time to realize that there was no machine learning in the paper. So I would say the findings are not clearly communicated.

W3. There also seems to be more significant issues where the evaluation score is built into the method in the simulated message routing study in Section 5. The method is scored by the same formula it optimizes. I have trouble understanding the justification for the empirical methodology.

Questions for authors

Can you clarify how these results advance science?

  1. what was previously unknown in the literature? what is the gap you identified?
  2. what is the theoretical or empirical evidence that you provide to fill this gap?

Comments on the paper

1 comment

Hi everyone, please acknowledge the authors’ rebuttal with a brief reply, even if your assessment is unchanged (if you already did or are about to do so, please ignore this message).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.