acceptodds
Under review as a conference paper at ICLR 2027

Routing Messages by What Teammates Know for Embodied Multi-Agent Cooperation

11 views

Abstract

Embodied agents built on large language models (LLMs) coordinate by exchanging natural-language messages, and each message must be delivered to some subset of teammates. Most existing methods choose these recipients from structure fixed in advance, such as roles, a communication graph, or a learned policy. However, how useful a message is to a teammate depends on what that teammate knows when the message is sent, and this changes at every step. We derive an approximation of a message's expected information gain for each recipient that depends on the recipient only through one fact: whether it already holds valid information about the goal-relevant entities the message mentions. Choosing the recipients of each message therefore reduces to tracking which agent knows about which entity, and the messages the agents exchange provide the evidence for this. We present CREAC (Cooperative Router for Embodied Agent Communication), a training-free router that infers from the team's messages which agents know about which entities, delivers each message only to the teammates for whom it is expected to be informative, and asks a teammate directly when it can no longer tell who knows. CREAC does not modify the agents' planners or prompts; it changes only which messages each agent receives. On household tasks with three to six agents and wildfire response with fifteen heterogeneous agents, CREAC outperforms broadcasting to all teammates in all ten combinations of planner, benchmark, and team size, reducing steps by up to 28% on household tasks and damage by 21–30% on wildfire response. The precision of its recipient selection increases with team size. In a user study, participants preferred teams that used CREAC over teams that broadcast every message and teams that did not communicate.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Multi Agent Cooperation, Multi Agent System, Large Language Model (LLMs), Embodied Agents, POMDP

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_Mo9EOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies the message routing problem in embodied multi-agent systems and proposes CREAC, a training-free dynamic communication routing method. Experiments on two embodied collaboration benchmarks, C-WAH and CREW-Wildfire, demonstrate that CREAC can reduce task execution costs and outperform various communication strategies and routing baselines.

Critical strengths
  1. The paper proposes a simple and clear message allocation strategy to address the problem of excessive irrelevant information in multi-agent communication.
  2. The experimental design is relatively comprehensive and effectively demonstrates that the performance improvements come from message selection.
  3. The method provides good interpretability, as each message selection decision is associated with an explicit rationale.
Critical weaknesses
  1. For some tasks, knowing which agent possesses certain information may not be sufficient to determine who should receive a message. This issue affects the core claim of the paper, namely that effective routing can be achieved by identifying which agents know what information. In large-scale tasks involving agents with different roles and capabilities, this assumption may be insufficient. Although the authors acknowledge in the appendix that action requests should theoretically consider whether an agent has the capability to complete the task, the actual implementation still relies on the simple information representation.
  2. CREAC only maintains a simple binary marker for each agent–entity pair, which may be too coarse to represent the complex knowledge states required in challenging tasks.
Questions for authors

The paper would benefit from a clearer overview figure illustrating the overall workflow of CREAC.

2 comments
AuthorsOpenReview ↗

Response to Reviewer Mo9E (1/2)

We thank the reviewer for the assessment of the method's clarity, interpretability and experimental design, and for the close reading of Appendix A.4. We have uploaded a revised PDF. Section, table and figure numbers below refer to the revised PDF. It adds an overview figure (Figure 1) and a TDW-MAT table (Table 3), so most later table and figure numbers are one higher than in the submitted version. Tables R1 and R2 are new results and keep their numbers from the responses to the other reviewers.

W1. Is knowing which agent holds which information sufficient when agents differ in roles and capabilities?

The paper evaluates CREAC in this setting on CREW-Wildfire, a large-scale task whose fifteen agents are of four types with different capabilities (Table 12). Drones can only scout. Helicopters carry firefighters and drop water. The bulldozer clears firebreaks. Firefighters suppress fire and rescue civilians. In this setting, CREAC gives the lowest damage of all communication policies in Table 27, under both planners. We also ran two graph-based routers, AgentPrune and GDesigner, and CREAC gives lower damage than both (Table R1).

Table R1 (Tables 2 and 30). CREW-Wildfire with fifteen agents of four types. Damage (lower is better) and tokens per episode, with the standard error over five seeds after ±.

CommunicationCoELA Damage ↓CoELA Usage ↓ReAct Damage ↓ReAct Usage ↓
Broadcast108.2 ± 7.85.55M112.4 ± 8.05.40M
RCR-Router104.8 ± 7.35.44M117.8 ± 8.55.21M
AgentPrune (new)94.3 ± 9.05.35M92.8 ± 7.35.17M
GDesigner (new)95.5 ± 8.35.14M97.1 ± 7.55.22M
CREAC76.2 ± 7.15.36M89.2 ± 7.55.15M

We also added a third benchmark, ThreeDWorld Multi-Agent Transport (TDW-MAT; Zhang et al., 2023), in which a team transports target objects to a goal location in multi-room 3D houses, with episodes of up to 3,000 steps. An agent carries at most two objects at a time, or more if it holds a container, so whether an agent can take on a transport depends on what it already carries. We ran its 24 official test episodes with teams of three to six agents and four planners, CoELA, ReAct, PCE (Seo et al., 2026) and Mind-Map Agent (Jeon et al., 2025). Overall, CREAC has the highest average transport rate under each of the four planners (Table R2).

Table R2 (Table 3). TDW-MAT. An episode lasts up to 3,000 steps, twelve times the horizon of C-WAH (250 steps) and fifteen times that of CREW-Wildfire (200 steps). An agent carries at most two objects at a time, or more with a container. Each cell gives the transport rate, the percentage of target objects transported (higher is better), averaged over teams of three to six agents, and in parentheses the mean tokens per episode, in millions and including the router. Each team size has one run of each of the 24 official test episodes. Bold marks the highest transport rate for each planner. Original is the communication that comes with PCE and Mind-Map Agent; like Broadcast, it sends every message to every teammate. The results for each team size, with standard errors, are in Table 28 (Appendix A.17).

PlannerCREACBroadcastRCR-RouterAgentPruneGDesigner
CoELA80.4 (0.83M)74.8 (0.76M)74.1 (0.79M)73.2 (0.79M)72.9 (0.79M)
ReAct86.1 (0.56M)79.5 (0.55M)82.9 (0.54M)82.4 (0.54M)82.3 (0.53M)
PlannerCREACOriginalRCR-RouterAgentPruneGDesigner
PCE95.7 (1.04M)92.3 (1.02M)93.8 (0.99M)94.0 (0.99M)93.3 (1.00M)
Mind-Map86.7 (1.11M)82.7 (1.04M)82.4 (1.02M)82.7 (0.99M)80.5 (1.03M)

In all of these results, CREAC routes requests for actions with the same meta-belief as requests for information, without a model of each agent's capabilities (Appendix A.4, Remark 1). In these teams of agents with different roles and capabilities, knowing which agent holds which information was therefore enough for CREAC to perform best among the compared methods.

AuthorsOpenReview ↗

Response to Reviewer Mo9E (2/2)

W2. Is a binary marker for each agent–entity pair too coarse?

The meta-belief is not binary. Each entry is Mt(i,e)=B(i,e) Kt(i,e)M_t(i,e) = B(i,e)\,K_t(i,e), the product of an evidence indicator and a confidence, and it takes continuous values between 0 and 1 (Section 4.3). The evidence indicator B(i,e)B(i,e) is 1 when the messages so far indicate that agent ii knows entity ee, and 0 otherwise. The confidence Kt(i,e)=exp⁡(−λ(t−ℓ(i,e)))K_t(i,e) = \exp(-\lambda (t - \ell(i,e))) decreases continuously with the number of steps since the entry was last updated. Routing compares Mt(i,e)M_t(i,e) with τrouting\tau_{\text{routing}}, and the router sends an active query when the largest entry for an entity falls below τquery\tau_{\text{query}}, so both decisions use the continuous value.

A meta-belief whose entries are binary corresponds to our ablation without decay (λ=0\lambda = 0), in which every entry stays at 0 or 1 (Section 6.5). On six-agent C-WAH, this ablation takes 51.85 steps, and CREAC takes 43.28 steps. The target precision of this ablation is 9.5 points lower than that of CREAC (Table 7).

In challenging tasks, the continuous meta-belief is sufficient for CREAC to outperform the other communication policies. On CREW-Wildfire, with fifteen agents of four types, CREAC gives the lowest damage of all communication policies under both planners (Table 27, and Table R1 in Part 1). On TDW-MAT, whose episodes last up to 3,000 steps, CREAC has the highest average transport rate under each of the four planners (Table R2 in Part 1).

We have revised Section 7 and Appendix A.22 of the revised PDF so that they describe each entry of the meta-belief as this continuous value.

Q. Overview figure

We agree, and the revised PDF adds an overview figure at the start of Section 4 (Figure 1).

We hope that these clarifications address the reviewer's concerns, and we would be glad to answer further questions. If they do, we would appreciate it if the reviewer would consider updating the assessment.

Reviewer_ju1EOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes CREAC, a training-free communication router for embodied multi-agent cooperation that dynamically selects message recipients based on what teammates are estimated to know. Motivated by expected information gain, CREAC maintains a meta-belief over agents' knowledge of goal-related entities, inferred from exchanged messages and updated through temporal decay. It routes information-sharing messages to agents likely to lack the relevant knowledge and requests to agents likely to possess it, while introducing active queries to refresh uncertain knowledge estimates. The method is evaluated on C-WAH with 3–6 agents and CREW-Wildfire with 15 heterogeneous agents. CREAC leading to more effective and efficient task solving, reducing steps by up to 28% on C-WAH and damage by 21–30% on CREW-Wildfire. Ablation studies further validate its routing and active querying mechanisms.

Critical strengths
  1. The paper study on an important problem of efficient message routing in multi-agent teams, aiming to deliver information to agents who need it and queries to those who can answer them, while avoiding unnecessary message delivery to other teammates.

  2. The ablation studies and controlled message-exposure experiments provide useful evidence for the effectiveness of selective routing and active querying.

Critical weaknesses
  1. The paper approximates the information gain of a message solely based on whether agents possess valid information about goal-related entities. This overlooks other important information such as task decomposition and allocation, agent locations and execution states, and future plan sharing. Such representation can not fully capture the actual value of a message for coordination and decision making.

  2. The experimental scale is limited, mainly use a single model backbone, covering only 10 distinct C-WAH episodes and five random seeds on CREW-Wildfire. The underlying planners choose are CoELA and ReAct with broadcast communication, while recent works such as PCE and Mind-Map Agent have demonstrated that reducing unnecessary messages can improve task efficiency. Without comparisons with these communication-efficient frameworks, it remains unclear whether CREAC still provides meaningful benefits when agents already communicate selectively, or whether much of the improvement could be achieved by suppressing unnecessary messages rather than routing them to specific recipients.

Questions for authors
  1. How much redundant communication actually occurs in the evaluated settings? Could the authors report the average number of messages sent and received per agent, as well as the proportion of messages filtered by CREAC? This would help clarify the practical significance of redundant message delivery.

  2. Why not allow agents to select message recipients directly? Like let each agent determines the recipients of its own messages when send the message, can the agent choose the correct agent to send to?

  3. Could the authors clarify and expand the evaluation scale? C-WAH includes only 10 distinct test episodes, while CREW-Wildfire reports results over five random seeds. How many independent scenarios or rollouts are evaluated in CREW-Wildfire? Would the improvements remain consistent across more diverse and challenging tasks?

7 comments
AuthorsOpenReview ↗

Response to Reviewer ju1E (1/5)

We thank the reviewer for the careful review. We have uploaded a revised PDF, which adds an overview figure of CREAC at the start of Section 4 (Figure 1). Section, table and figure numbers below refer to the revised PDF. Because of the new Figure 1 and the new Table 3 (TDW-MAT), most later table and figure numbers are one higher than in the submitted version. Tables R2 to R6 are new results. Table R2 is from a third benchmark, ThreeDWorld Multi-Agent Transport (TDW-MAT; Zhang et al., 2023), in which a team transports target objects to a goal location in multi-room 3D houses. Tables that share a number come from the same experiment; each response shows the part that its questions concern.

W1. Does representing messages by goal-related entities miss task allocation, agent locations, execution states and plans?

To use this information in estimating the value of a message for each recipient, the router would need to know, for each recipient, which task it has been allocated, where it is, what it is doing and what it plans to do next. This information is held inside each agent's planner, and the router sees only the messages that the agents send. Locations and execution states change at every step, but agents mention them in messages only occasionally. For example, if Alice says that she will go to the kitchen and then changes her plan, the router cannot know this until Alice says so in a later message. For the router to keep this information current, the planners would have to be changed so that agents report it at every step. CREAC is designed to be added to existing planners without changing them. A router that tries to read all of this information from the agents' messages, direct reasoning in Section 6.3, takes more steps and uses more tokens than CREAC. We explain these two points in turn.

(1) Why obtaining this information requires modifying the planners. The router cannot read the internal states of the agents, because that would break the DEC-POMDP-COM setting of the paper (Section 4.2). This setting matters in practice. In a team that includes humans, or agents built by different developers on different language models, a router cannot read every member's internal state; it sees only what the members say. A router that reads internal states would work only when every member exposes its internals, whereas a router that reads only messages, as CREAC does, can also serve such teams (Appendix A.2 of the revised PDF). What an agent knows about a goal-related entity can be inferred from every message that mentions the entity. Under the cooperative premise of Section 3, each SHARE and each QUERY indicates whether its sender and its recipients hold information about the entities that it mentions, and rules (R1) and (R2) update the meta-belief from it (Section 4.2). An agent's location, execution state and plan cannot be inferred in this way. They change from step to step without any message, and the router learns them only when the agent's planner decides to state them. The router cannot control when or in what form a planner states its plan, so a plan that an agent stated earlier can change without any message that the router could read. With CoELA, ReAct, PCE and Mind-Map Agent, an agent sends between 2.35 and 3.89 messages per C-WAH episode (Table R5 in Part 3). For the router to have this information at every step, either every message would have to include it, which changes what the planner writes, or the agents would have to send messages more often, which changes when the planner speaks. Both are modifications of the planner. CREAC reads only the messages that the planners already send, and this is why it can be added to CoELA, ReAct, PCE and Mind-Map Agent without changing them (Tables 1 and 3).

AuthorsOpenReview ↗

Response to Reviewer ju1E (2/5)

(2) A router that tries to read all of this information from the messages. Direct reasoning (Section 6.3) is a router in which a language model chooses the recipients of each message instead of the meta-belief. We analyzed the reasoning traces of gpt-oss-20b and found that the model tries to infer each agent's task allocation, location, execution state and plan from the dialogue history, and to use them to choose the recipients and to decide whether to send a query. The following is an example. At step 20 of a six-agent C-WAH episode, Eve sends a message that reassigns the goal items, and gpt-oss-20b reasons: "Bob's earlier message said: "Alice – give me pudding 373 and juice 369." So Bob expected to get them. But Eve's message: "Alice – hand me pudding 373 & juice 369." [...] Bob still may want to know that the plan is to hand to Eve. But Bob is not in recipients unless it says he needs to take action. [...] All have been found and their positions known [...] the table is living-room. [...] Thus no query required." The model infers Bob's allocation and plan from his earlier message, and the progress of the task and the positions of the items from other messages. It then sends Eve's message to Alice, Carol, David and Frank, and not to Bob, whose plan the message changes.

The table below gives the results on six-agent C-WAH from Table 5, with the standard errors over five seeds after ±.

RouterSteps (lower is better)Tokens per episode
Direct reasoning, gpt-oss-20b61.02 ± 4.6124k ± 10k
Direct reasoning, GPT-555.63 ± 4.2130k ± 11k
CREAC43.28 ± 2.085k ± 5k

With both language models, direct reasoning takes more steps and uses more tokens than CREAC.

W2 and Q3. Scale of the evaluation

Language model. The reviewer's question can refer to the language model of the agents or to the language model of the router, and we answer both. The agents run on gpt-oss-20b in all experiments, including TDW-MAT, except where stated otherwise. For the agents, Appendix A.8 repeats the six-agent C-WAH experiment with Qwen3-32B agents, and CREAC takes 46.8 steps while Broadcast takes 49.9 steps. For the router, which uses a language model to classify each message and extract its entities, we ran Qwen3-32B in place of gpt-oss-20b (Table R3). With either language model in the router, CREAC outperforms Broadcast in all ten combinations of benchmark, planner and team size.

Table R3 (Table 31). Qwen3-32B as the router's language model, with the agents on gpt-oss-20b. C-WAH: steps (lower is better), with the standard error over five seeds of ten episodes each. CREW-Wildfire: damage (lower is better), with the standard error over five seeds.

PlannerRouterC-WAH 3C-WAH 4C-WAH 5C-WAH 6CREW 15
CoELABroadcast62.26 ± 2.458.83 ± 2.250.63 ± 2.151.33 ± 2.3108.2 ± 7.8
CREAC, gpt-oss-20b55.45 ± 1.948.27 ± 1.944.90 ± 2.043.28 ± 2.076.2 ± 7.1
CREAC, Qwen3-32B54.17 ± 1.748.43 ± 2.143.43 ± 2.147.07 ± 2.181.3 ± 7.3
ReActBroadcast76.32 ± 3.673.72 ± 3.469.65 ± 3.169.33 ± 3.2112.4 ± 8.0
CREAC, gpt-oss-20b61.40 ± 2.255.70 ± 2.451.30 ± 2.249.76 ± 2.389.2 ± 7.5
CREAC, Qwen3-32B59.31 ± 1.653.18 ± 2.251.75 ± 2.348.74 ± 1.892.5 ± 7.8

Episodes and rollouts. The ten C-WAH episodes are the complete official test set of the benchmark, which prior work also uses (Zhang et al., 2023). We run each episode with five seeds, so each cell averages 50 runs. Each CREW-Wildfire configuration has five rollouts, one per seed, of a 200-step episode on a procedurally generated map with stochastic fire spread (Appendix A.5). Every policy runs on the same five seeds.

A third benchmark, TDW-MAT. To evaluate CREAC on more diverse and longer tasks, we added TDW-MAT. Its 24 official test episodes cover two task types, food and stuff, in multi-room 3D houses, with a horizon of 3,000 steps. An agent carries at most two objects at a time and must find and use a container to carry more. We ran teams of three to six agents with four planners and five communication policies, one run per episode, which gives 80 configurations and 1,920 episodes.

AuthorsOpenReview ↗

Response to Reviewer ju1E (3/5)

Table R2 (Table 3). TDW-MAT. An episode lasts up to 3,000 steps, twelve times the horizon of C-WAH (250 steps) and fifteen times that of CREW-Wildfire (200 steps). An agent carries at most two objects at a time, or more with a container. Each cell gives the transport rate, the percentage of target objects transported (higher is better), averaged over teams of three to six agents, and in parentheses the mean tokens per episode, in millions and including the router. Each team size has one run of each of the 24 official test episodes. Bold marks the highest transport rate for each planner. Original is the communication that comes with PCE and Mind-Map Agent; like Broadcast, it sends every message to every teammate. The results for each team size, with standard errors, are in Table 28 (Appendix A.17).

PlannerCREACBroadcast
CoELA80.4 (0.83M)74.8 (0.76M)
ReAct86.1 (0.56M)79.5 (0.55M)
PlannerCREACOriginal
PCE95.7 (1.04M)92.3 (1.02M)
Mind-Map86.7 (1.11M)82.7 (1.04M)

Overall, CREAC has a higher average transport rate than Broadcast or Original under each of the four planners. The horizon of TDW-MAT, 3,000 steps, is twelve times that of C-WAH, so more meta-belief entries decay below the query threshold during an episode. Active queries are the queries that the router sends by itself when the largest meta-belief entry for an entity is below the query threshold (Section 4.3). The router therefore sends more active queries than on C-WAH, and each active query adds a reply from an agent. As a result, CREAC mostly uses more tokens than Broadcast or Original on TDW-MAT, but the median difference over planners and team sizes is low, about 3%.

PCE and Mind-Map Agent. CREAC improves both planners, and PCE with CREAC is the best system on C-WAH and on TDW-MAT. We ran each planner with its own communication (Original), which sends every message to every teammate, and with CREAC's delivery, on C-WAH (Table R4) and on TDW-MAT (Table R2).

Table R4 (Tables 1 and 29). PCE (Seo et al., 2026) and Mind-Map Agent (Jeon et al., 2025) on C-WAH. Steps (lower is better), with the standard error over five seeds of ten episodes each. Tokens per episode, including the router, are in parentheses. Original is each planner's own communication.

PlannerCommunication3 agents4 agents5 agents6 agents
PCEOriginal44.90 ± 2.2 (383k)37.88 ± 1.7 (465k)34.66 ± 1.7 (546k)32.66 ± 1.6 (639k)
CREAC43.30 ± 1.9 (392k)37.40 ± 1.5 (466k)33.02 ± 1.8 (539k)29.56 ± 1.6 (592k)
Mind-MapOriginal53.86 ± 2.2 (225k)49.62 ± 1.7 (304k)41.52 ± 2.0 (333k)39.48 ± 1.7 (408k)
CREAC52.62 ± 1.5 (244k)46.78 ± 2.0 (308k)39.96 ± 1.6 (361k)36.58 ± 1.6 (423k)

On C-WAH, CREAC takes fewer steps than Original in all 8 settings, without any change to the planners. With six agents, it reduces the steps of PCE by 9.5% (32.66 to 29.56) and those of Mind-Map Agent by 7.3% (39.48 to 36.58). On TDW-MAT, CREAC gives a higher transport rate in 7 of the 8 settings and the same rate in the eighth. PCE with CREAC takes the fewest steps of all systems on C-WAH at every team size, and it has the highest average transport rate of all systems on TDW-MAT, 95.7%.

The planner decides how often an agent speaks: under Broadcast with CoELA and ReAct and under Original with PCE and Mind-Map Agent, an agent sends 2.35 to 3.89 messages per C-WAH episode at every team size (Table R5 in Part 3). The number of messages that each agent receives grows with team size, because every message reaches every teammate. With PCE and Original, an agent receives 6.27 messages per episode in a team of three and 15.77 in a team of six; with PCE and CREAC, it receives 8.61 in a team of six. CREAC_random, a control in the paper (named Random in the submitted version), sends each message to as many teammates as CREAC would select, chosen at random. On C-WAH, CREAC_random takes 5.95 to 10.82 more steps than CREAC under CoELA and ReAct (Tables 1 and 14), so the gain depends on which teammates receive each message.

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

This paper studies the message routing problem in embodied multi-agent cooperation, focusing on how to determine which agents should receive a given message. Existing approaches typically rely on broadcasting, predefined communication topologies, agent roles, or planner-coupled communication strategies, which may fail to adapt to what each agent currently knows. The authors aim to develop a knowledge-aware message routing mechanism that reduces redundant communication and improves multi-agent collaboration.

To achieve this goal, the authors propose CREAC, a training-free and planner-agnostic message routing framework. The key idea is to approximate the Expected Information Gain (EIG) of a message and maintain an Agent-Entity meta-belief matrix that tracks which agents possess valid information about task-relevant entities. Specifically, CREAC extracts knowledge-related information from natural-language messages, routes SHARE messages to agents that lack the relevant information, and routes QUERY messages to agents likely to have it. The framework also incorporates confidence decay and active querying to handle outdated or missing knowledge, without requiring modifications to the agents' planners.

The authors evaluate CREAC on two benchmarks: C-WAH, involving 3–6 agents performing household tasks, and CREW-Wildfire, involving 15 heterogeneous agents in wildfire response scenarios. Experiments are conducted with two planners, CoELA and ReAct, and compared against multiple baselines, including Broadcast, Random, RCR-Router, CoTS, and REVECA. The results show that CREAC consistently outperforms broadcasting across all ten evaluated configurations, reducing task completion steps by up to 28% on C-WAH and damage cost by 21–30% on CREW-Wildfire. Additional ablation studies and a user study further support the effectiveness of selective routing and active querying.

Critical strengths
  • The paper addresses an important and practical problem. Redundant communication can be a serious issue in embodied multi-agent systems, especially as the number of agents increases.

  • The proposed framework is simple and flexible. CREAC is training-free and planner-agnostic, making it easy to integrate into existing multi-agent systems without modifying their planners.

  • The experimental evaluation is fairly comprehensive. The authors evaluate CREAC on two benchmarks with different team sizes and planners, showing consistent improvements over broadcasting. The ablation studies cover several key components, and the message exposure experiment provides additional evidence of why redundant communication can hurt agent performance. The authors also conduct a human study, where participants generally prefer CREAC over broadcasting and no communication.

Critical weaknesses
  • The technical novelty seems somewhat limited. The idea of tracking what each agent knows is closely related to existing work on transactive memory and knowledge-aware communication. The meta-belief representation, confidence decay, and threshold-based routing are also relatively simple mechanisms. The EIG formulation relies on strong simplifying assumptions, which makes the method less interesting. I would like to see a clearer discussion of what is new compared to existing knowledge-aware communication methods.

  • The comparison with existing methods could be stronger. Most baselines in Table 2 are relatively simple heuristics, such as Random, Recency, and Distance. RCR-Router is the main existing routing method, but the authors use an adapted implementation, making it unclear whether the comparison fully reflects its original capabilities. Based on the provided numbers in Table 1 & 2, RCR-Router seems about the same as Random. Including stronger communication baselines or providing more justification for the baseline adaptations would make the experimental results more convincing.

Questions for authors
  • In Tables 1 and 2, RCR-Router performs similarly to Random, which is somewhat surprising and not intuitive given its reported performance in the original paper. Could the authors explain why this happens and whether the adaptation affects its performance?

  • Could the authors clarify what the contribution of EIG formulation in practice? Personally, I feel the actual implementations fall into a series of heuristic and rule-based decisions.

  • In Table 5, Broadcast+ActiveQuery delivers messages to all agents, but the meta-belief is only updated for recipients selected by CREAC (Appendix A.14). Would the results change if the meta-belief were updated based on the actual broadcast deliveries? I feel this design may put Broadcast at a disadvantage, as it prevents the router from fully accounting for the information that agents actually receive.

9 comments
AuthorsOpenReview ↗

Response to Reviewer rXHo (1/4)

We thank the reviewer for the detailed review and the questions. We have uploaded a revised PDF. Section, table and figure numbers below refer to the revised PDF. It adds an overview figure (Figure 1) and a TDW-MAT table (Table 3), so most later table and figure numbers are one higher than in the submitted version. Tables R1 and R2 are new results. Table R2 is from a third benchmark, ThreeDWorld Multi-Agent Transport (TDW-MAT; Zhang et al., 2023), in which a team transports target objects to a goal location in multi-room 3D houses.

W1(a). What is new relative to transactive memory and knowledge-aware communication

CREAC decides, for each message during an episode, which teammates receive it, from an estimate of what each teammate knows at that step. We have revised the extended related work in Appendix A.1, which now compares CREAC with each line of work below, and we summarize the differences here. The related methods decide communication from other information, and each of them needs something that CREAC does not need. Decision-theoretic methods (Roth et al., 2005; Becker et al., 2009) compute the value of a message from a full model of the team and from policies computed before execution. Learned targeted communication, such as TarMAC (Das et al., 2019), I2C (Ding et al., 2020) and Who2com (Liu et al., 2020), requires training on the task. CoThinker (Shang et al., 2025) gives each agent the outputs of peers chosen by embedding similarity, and its transactive memory records each agent's expertise. Multi-Agent Transactive Memory (Kim et al., 2026) stores agents' trajectories so that other agents retrieve them in later tasks. Li et al. (2023) keep a belief state in each agent's prompt, and every message still goes to all teammates. AgentPrune (Zhang et al., 2024a), GDesigner (Zhang et al., 2024b) and MARBLE (Zhu et al., 2025) use communication graphs or topologies that are optimized, trained or set before execution, and RCR-Router (Liu et al., 2025) selects context by role, task stage and token budget. CREAC needs no environment model, no training and no change to the agents; it uses only the messages that the agents send and the goal description of the task.

W1(b). Simplicity of the mechanisms

The simple mechanisms make CREAC flexible, training-free and planner-agnostic, so it can be integrated into an existing multi-agent system without modifying its planners. Because CREAC does not modify the planners, we added it in this rebuttal to two further planners, PCE (Seo et al., 2026) and Mind-Map Agent (Jeon et al., 2025), without any change (Appendix A.18). Its hyperparameters were set once on three-agent C-WAH and used unchanged for every team size and benchmark. Each routing decision can be traced to the meta-belief entries that produced it.

W1(c). The assumptions behind the EIG approximation

The assumptions are the standard ones for cooperative multi-agent teams. We did not simplify the benchmarks, and we did not modify them to fit our assumptions. Instead, we also run experiments with larger teams. For example, C-WAH originally has two agents, and we run it with three to six. The paper states the assumptions to show how a planner-agnostic routing rule follows from the expected information gain (EIG) of a message.

The setting. We model the team as a DEC-POMDP-COM in which the agents share one goal, and we assume that agents do not withhold or fabricate information (Section 3). C-WAH, CREW-Wildfire and TDW-MAT are fully cooperative, and CoELA, CoTS and REVECA, which we compare against, assume the same setting.

Assumptions (A1) to (A3), which the paper tests. (A1) states that a message has different value for different recipients. CREAC's target precision exceeds that of delivering to every teammate by 14 to 38 points (Table 4). (A2) states that a message with low expected gain makes its recipient act worse. At 500 decision points of six-agent C-WAH, adding to an agent's context eight messages that CREAC had withheld from it lowers the share of productive actions by 10.4 points (Section 6.4, Figure 2). (A3) states that better individual actions improve the result of the team, which Tables 1 and 2 show through team performance. Of 300 messages that we labeled by hand, 298 mention a goal-related entity (Table 23), so restricting the gain to goal-related entities leaves out almost no messages.

Why the result may interest the ICLR community. With the planner, the prompts and the agents' language model fixed, CREAC changes only which messages each agent receives. This change alone reduces the number of steps by up to 28% on C-WAH and the damage by up to 30% on CREW-Wildfire, compared with Broadcast (Tables 1 and 2).

AuthorsOpenReview ↗

Response to Reviewer rXHo (2/4)

W2. Stronger baselines

We added two published graph-based routers, AgentPrune and GDesigner, which the paper compares on six-agent C-WAH (Table 13). We ran them on CREW-Wildfire (Table R1) and, together with RCR-Router, on TDW-MAT (Table R2). These three methods were evaluated on question answering and coding, and we adapted them to embodied teams as described in Appendix A.14. Each change gives the method more information or more flexibility than the original. The paper's other baselines are the planner-coupled systems CoTS and REVECA, direct reasoning with gpt-oss-20b and GPT-5, and four controls that each change one part of the communication policy.

Table R1 (Tables 2 and 30). CREW-Wildfire with fifteen agents of four types. Damage (lower is better) and tokens per episode, with the standard error over five seeds after ±.

CommunicationCoELA Damage ↓CoELA Usage ↓ReAct Damage ↓ReAct Usage ↓
Broadcast108.2 ± 7.85.55M112.4 ± 8.05.40M
RCR-Router104.8 ± 7.35.44M117.8 ± 8.55.21M
AgentPrune (new)94.3 ± 9.05.35M92.8 ± 7.35.17M
GDesigner (new)95.5 ± 8.35.14M97.1 ± 7.55.22M
CREAC76.2 ± 7.15.36M89.2 ± 7.55.15M

On CREW-Wildfire, AgentPrune and GDesigner give 12–17% lower damage than Broadcast, and CREAC gives lower damage than both under both planners.

Table R2 (Table 3). TDW-MAT. An episode lasts up to 3,000 steps, twelve times the horizon of C-WAH (250 steps) and fifteen times that of CREW-Wildfire (200 steps). An agent carries at most two objects at a time, or more with a container. Each cell gives the transport rate, the percentage of target objects transported (higher is better), averaged over teams of three to six agents, and in parentheses the mean tokens per episode, in millions and including the router. Each team size has one run of each of the 24 official test episodes. Bold marks the highest transport rate for each planner. The results for each team size, with standard errors, are in Table 28 (Appendix A.17).

PlannerCREACBroadcastRCR-RouterAgentPruneGDesigner
CoELA80.4 (0.83M)74.8 (0.76M)74.1 (0.79M)73.2 (0.79M)72.9 (0.79M)
ReAct86.1 (0.56M)79.5 (0.55M)82.9 (0.54M)82.4 (0.54M)82.3 (0.53M)

Overall, CREAC has the highest average transport rate under both planners.

Gains and token cost. On CREW-Wildfire, CREAC uses the fewest tokens with ReAct and at most 4% more tokens than any other method with CoELA, and it gives 19% (CoELA) and 4% (ReAct) lower damage than the best other method (Table R1). On TDW-MAT, the benchmark with the longest horizon, CREAC gives the highest average transport rate under every planner (Table R2). Previous gains on TDW-MAT came from new agent designs, such as CoELA (Zhang et al., 2023) and the Mind-Map Agent (Jeon et al., 2025). CREAC adds its gain on top of each planner without changing it, including the two most recent planners, PCE and the Mind-Map Agent (Table 3 and Appendix A.17). With PCE, CREAC reduces the share of target objects left untransported from 7.7% to 4.3%. For this gain, CREAC uses a median of about 3% more tokens than each planner's own communication over planners and team sizes. We attribute the extra tokens on TDW-MAT to active queries, the queries that the router sends by itself when no teammate is likely to hold valid information about an entity (Section 4.3). The horizon of TDW-MAT, 3,000 steps, is twelve times that of C-WAH, so more meta-belief entries decay below the query threshold during an episode, and each active query adds a reply from an agent.

AuthorsOpenReview ↗

Response to Reviewer rXHo (3/4)

Q1. Why does RCR-Router perform similarly to Random, and does the adaptation affect its performance?

Random is a control, not a naive baseline. It sends each message to as many teammates as CREAC selects and keeps CREAC's active queries, so it differs from CREAC only in which teammates receive each message (Appendix A.14). Because the name Random suggests a naive baseline, we renamed it CREAC_random in the revised PDF.

Why we compare with RCR-Router. To our knowledge, no earlier method selects the recipients of each message in an embodied LLM team during execution without changing the agents' planner. RCR-Router is the closest published method: in each round, it selects the shared-memory items that each agent reads.

Why RCR-Router's original setting differs. RCR-Router was designed for and evaluated on multi-hop question answering, where each agent has a predefined role, such as Planner or Searcher, and the information that an agent needs follows from its role (Liu et al., 2025). It was not designed for long-horizon, partially observable tasks. Our agents have no predefined information roles, so our adapter infers each agent's role from the entities that it mentioned recently and scores recipients by this role, the task stage and recency (Appendix A.14). In our benchmarks, however, the information that an agent needs depends on what it has observed, which changes at every step. In CREW-Wildfire, for example, ten of the fifteen agents are firefighters, and whether a firefighter lacks the location of a new fire depends on whether it has observed the fire, not on its type or on the entities that it mentioned recently. Role, task stage and recency do not record what each agent has observed or whether its information is still valid, and this matters more as the horizon grows.

Q2. What does the EIG formulation contribute in practice?

If the EIG formulation contributed nothing, the rules that it gives would be just one choice among many reasonable heuristics, and other reasonable rules should work as well. We therefore replaced each part of the implementation that follows from the formulation with a natural alternative that a designer might choose without the formulation. Every alternative performs worse: on six-agent C-WAH with CoELA, each takes 17–38% more steps than CREAC (43.28 steps).

Part of the formulationAlternativeWhy the alternative is naturalSteps ↓
Information becomes invalid over time, so the confidence decays (Appendix A.3)No decay (λ=0\lambda = 0)A fact that an agent has received seems to stay known51.85 (Table 7)
A delivered SHARE gives the recipient valid information, so its entry is set to 1 (Section 4.2)The entry set to 0.5A recipient may not use every message50.5 (Table 20)
A SHARE has gain for agents that lack its information, and a QUERY for agents that hold it (Equations 2 and 3)The same number of recipients, chosen at randomIt removes as many deliveries as CREAC without tracking who knows what54.11 (Table 6)
A message has no gain for an agent that already knows its content (Equation 1)Every message to every agent, with the same active queriesIt is the planners' standard channel58.30 (Table 6)
A query has the most value when no agent has evidence about the entity (Appendix A.3)No query in that caseNo agent is known to have the answer59.6 (Table 21)
The agent whose entry was updated longest ago most likely holds unreported information (Appendix A.3)The query sent to every agentAny agent may know the answer56.8 (Table 21)

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary for the Area Chair

Dear Area Chair,

Thank you for handling our submission. Below we summarize the reviews, our responses, and what still awaits a reply. Numbers refer to the revised PDF.

1. Strengths noted by the reviewers

  • Problem: routing each message to the agents who need it is important and practical, especially as teams grow (rXHo, ju1E).
  • Method: CREAC is simple, training-free and planner-agnostic, so it can be added to existing planners without changes (rXHo, Mo9E).
  • Experiments: CREAC improves consistently over broadcasting across benchmarks, team sizes and planners, and ablations show that the gains come from message selection (rXHo, ju1E, Mo9E).
  • Interpretability: each routing decision has an explicit rationale (Mo9E).

2. Concerns and our responses

Reviewer rXHo

  1. Baselines; RCR-Router performs like Random. We added two routers, AgentPrune and GDesigner, and a third benchmark, TDW-MAT (3D transport, up to 3,000 steps). CREAC is the best method on CREW-Wildfire and TDW-MAT under every planner (Tables 2 and 3). Random (renamed CREAC_random) is a control that differs from CREAC only in who receives each message. RCR-Router scores recipients by role, task stage and recency, which do not record what each agent has observed. Reviewer rXHo acknowledged this point.
  2. Novelty and the contribution of the EIG formulation. Prior methods track slow-changing expertise or roles; CREAC tracks whether each agent's information about each entity is still valid during an episode, without training or changes to the agents (Appendix A.1). Replacing each formulation-derived part with a natural alternative costs 17–38% more steps on six-agent C-WAH. Reviewer rXHo acknowledged that these ablations show the components are effective. Our 16 Nov follow-up adds a decision that differs from natural heuristics: when no agent has evidence about an entity, CREAC queries the agent with the oldest entry (43.28 steps), against 59.6 for suppressing the query and 56.8 for broadcasting it (Table 21).
  3. Broadcast+ActiveQuery updates the meta-belief only for CREAC's recipients. Reviewer rXHo acknowledged the motivation of this controlled comparison. In our 16 Nov follow-up, updating the meta-belief on all broadcast deliveries took 62.09 steps, against 58.30 for our design, so our design did not disadvantage Broadcast.

Reviewer ju1E

  1. Evaluation scale. With Qwen3-32B as the router's model, CREAC outperforms Broadcast in all ten settings (Table 31); it also does so with Qwen3-32B agents (Appendix A.8). The ten C-WAH episodes are the full official test set, each run with five seeds. TDW-MAT adds 1,920 episodes (3–6 agents, four planners).
  2. No comparison with PCE and Mind-Map Agent; gains may come from suppression alone. Added to both planners without changes, CREAC takes fewer steps in all 8 C-WAH settings and gives a higher or equal transport rate in all 8 TDW-MAT settings. PCE with CREAC is the best system on both benchmarks (Tables 1 and 3). CREAC_random, which removes as many deliveries as CREAC but at random, takes 5.95–10.82 more steps, so the gain depends on who receives each message.
  3. Entities miss task allocation, locations, states and plans. This information is held inside each planner, and the router sees only messages. When GPT-5 infers this information from the dialogue to choose recipients, the team takes 55.63 steps and 130k tokens, against 43.28 and 85k with CREAC (Table 5).
  4. Amount of redundant communication; letting agents choose recipients. With six agents on C-WAH, CREAC reduces the deliveries that each agent receives by 28–45% compared with broadcasting (Table 32). Agents can already choose recipients: CREAC always delivers to teammates the sender names.

Reviewer Mo9E

  1. Knowing who knows what may be insufficient when agents differ in roles and capabilities. CREW-Wildfire has four agent types with different capabilities, and TDW-MAT limits how many objects an agent can carry. Without a capability model, CREAC is the best method on both benchmarks under every planner.
  2. A binary agent–entity marker is too coarse. Each entry is continuous: an evidence indicator times a confidence that decays over time. A binary version (no decay) takes 51.85 steps, against 43.28 for CREAC (Table 7).
  3. Overview figure. Added as Figure 1.

Reviewer Mo9E acknowledged our response and has no further questions.

3. Responses awaiting a reply

  • ju1E: We answered every point (items 4–7) on 14 Nov and followed up on 18 Nov, but have not yet received a reply.
  • rXHo: Our 16 Nov follow-up ran the requested Broadcast+ActiveQuery variant (item 3) and explained what the EIG formulation adds (item 2), but has not yet received a reply.

We hope these responses can be considered in the AC–reviewer discussion.

Thank you again for your time and effort.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.