acceptodds
Under review as a conference paper at ICLR 2027

Enhancing Multi-Agent Communication through Attention Steering with Context Relevance

Abstract

LLM-based multi-agent systems often accumulate long communication histories. As a result, relevant information is increasingly diluted by irrelevant context, leading to degraded performance. Existing context management methods compress or prune the history, which may discard subtle but critical information. Attention steering preserves the full history, but identifying relevant context remains challenging as multi-agent communication evolves at every reasoning step. To address this, we introduce Agent-Radar, an attention steering method that leverages the inherent structure of multi-agent communication to dynamically guide agents’ attention toward relevant context. Across five benchmarks, Agent-Radar outperforms state-of-the-art methods, yielding average gains of 4.17 absolute points over the strongest baseline. Furthermore, our analysis shows that Agent-Radar remains effective and robust as the number of agents and interaction rounds increases. Finally, the ablations show that the core components in Agent-Radar are crucial to performance.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Multi-Agent System, Attention Steering, Agent Communication, Long-Context Reasoning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 4 reviews, above 85% of the others.
Reviewer_31M5OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces Agent-Radar, a training-free method for helping LLM agents use relevant information in accumulated communication histories. It selects sentences based on semantic similarity to the current query, weighted by message age and the graph distance between the source and receiving agents. The selected context is then emphasized using an existing attention-steering method while retaining the full history. Experiments across five benchmarks, multiple multi-agent frameworks, and different communication topologies show consistent improvements.

Critical strengths
  1. The empirical evaluation is fairly thorough. Agent-Radar improves performance across five benchmarks and three multi-agent frameworks, with additional experiments on different topologies and backbone models.

  2. The method is simple and the ablations help explain its benefits. The paper compares against semantic-only retrieval, random steering, alternative retrievers, and a prompt-only retrieval variant, and separately removes spatial and temporal decay. Together, these experiments provide useful evidence that both context selection and steering contribute to performance. The method also preserves the existing workflow and does not require model training.

Critical weaknesses
  1. The methodological contribution is relatively modest. The steering mechanism is inherited from SPA, while the main addition is a sentence-selection score combining semantic similarity with spatial and temporal decay. The results support the usefulness of this combination, but the paper should distinguish its contribution to automatic context selection more clearly from advances in attention steering itself.

  2. The assumptions behind the decay factors are not directly tested. Recent messages and nearby agents may often be useful, but important evidence can also come from an earlier round or a distant specialist. This is particularly relevant when the communication graph is randomly constructed, since graph distance need not reflect information quality. The component ablation demonstrate average benefits, but do not establish when these preferences help or suppress useful evidence.

Questions for authors
  1. How does Agent-Radar behave when decisive evidence appears in an older message or comes from a distant agent, while recent or nearby messages contain incorrect reasoning? A controlled experiment would help establish the limits of the decay assumptions.

  2. In the spatial and temporal ablations, does removing a decay factor substantially change the number of selected sentences? A comparison with the selection count held fixed would help separate the benefit of the relevance ranking from changes in how much context is emphasized.

  3. Can you clarify whether the prompt-token accounting in the efficiency analysis includes the masked-input computation used by SPA? A breakdown of retrieval and steering latency, together with the context lengths used, would make the reported overhead easier to interpret.

3 comments
AuthorsOpenReview ↗

Response to Reviewer 31M5 (1/2)

Dear Reviewer 31M5,

Thank you for the careful review. We address each point below.

W1/Q1: The paper should more clearly distinguish its contribution to automatic context selection from the underlying attention steering mechanism (SPA).

Thank you for this comment. We would like to clarify that SPA [1] requires users to manually specify which input tokens the model should attend to. In the original SPA experiments, the steered context is selected by human and fixed in advance. Given the complexity of multi-agent systems, the communication history is long and changes at every step, so it is hard to decide manually which parts to steer attention to.

By contrast, Agent-Radar automatically identifies the relevant context in the multi-agent conversation history. More importantly, Agent-Radar is agnostic to the steering method. We have conducted an experiment in Appendix A.5 replacing SPA with PASTA and show the same trend. We have updated the manuscript to clarify this point (Lines 49–52).

[1] Tian, Yuan, and Tianyi Zhang. "Selective prompt anchoring for code generation." ICML 2025.

[2] Zhang, Qingru, et al. "Tell your model where to attend: Post-hoc attention steering for llms." ICLR 2024.

W2/Q1 The decay assumptions are not directly tested, particularly for randomly constructed communication graphs. How does the method behave when decisive evidence appears exclusively in an older or distant message?

Thank you for raising this question. First, we would like to clarify that our use of randomly constructed communication topologies follows established prior work on the same topic, including AgentPrune [1] (ICLR 2025) and AgentDropout [2] (ACL 2025). Both studies investigate communication efficiency in multi-agent systems and adopt random communication topologies in their evaluations. Our experimental setup follows these established practices to ensure consistency and comparability with prior research.

To further investigate the concern about important information originating from earlier messages or more distant agents, we manually inspected 100 randomly sampled tasks. We found that only 9% of the tasks contained critical information exclusively in earlier messages or messages from more distant agents, without being restated in subsequent communication.

To understand how Agent-Radar behaves in these cases, we compare it with the original multi-agent system without steering and Agent-Radar with only semantic-based steering (no decay). We evaluated all three methods on both the 9 edge cases and the full set of 100 tasks across three runs.

Method9 edge casesAll 100 tasks
No steering3.33 ± 0.5861.67 ± 1.53
semantic-based steering6.67 ± 0.5864.00 ± 2.65
Agent-Radar6.00 ± 1.0070.33 ± 1.53

As expected, semantic-based steering performs best on the 9 edge cases (6.7 correct answers on average), as it can directly steer attention toward relevant information regardless of its temporal or spatial distance. However, it does not achieve the best overall performance across all 100 tasks, since relying solely on semantic similarity without considering recency and spatial relationship may introduce indirect noise, redundant context, or outdated information. Agent-Radar achieves the highest overall performance (70.3 correct answers on average, compared with 64.0 for semantic-based steering). Although Agent-Radar does not explicitly steer attention toward the distant critical information in the 9 edge cases, that information remains in the context window. Moreover, Agent-Radar still outperforms the original multi-agent system on the 9 edge cases (6.0 versus 3.3 correct answers on average), potentially because it explicitly steers attention toward the original question (Lines 211-212), preventing the task objective from being buried in the conversation history. We have added this analysis and discussion to Appendix C.

[1] Zhang, Guibin, et al. "Cut the crap: An economical communication pipeline for llm-based multi-agent systems." ICLR 2025.

[2] Wang, Zhexuan, et al. "Agentdropout: Dynamic agent elimination for token-efficient and high-performance llm-based multi-agent collaboration." ACL 2025.

AuthorsOpenReview ↗

Response to Reviewer 31M5 (2/2)

Q2 In the spatial and temporal ablations, does removing a decay factor substantially change the number of selected sentences?

Thank you for pointing this out. We would like to clarify that the "removing decay" ablations in Table 5 already hold the selection count fixed to isolate the benefit of the relevance ranking. As mentioned in Line 993, removing a decay factor will change the number of sentences receiving attention steering. To isolate each decay contribution, we match the steering budget to the full AGENT-RADAR method. Specifically, if Agent-Radar steers kk historical sentences, each ablation steers its top-kk sentences under the ablated scoring function. We have added a footnote on page 8 to make this experimental control more explicit.

Q3 Clarify the latency overhead breakdown

Thank you for this comment. We break down the end-to-end latency (seconds per task) into the table below:

ComponentBase MASAgent-Radar
Sentence retrieval–1.2
Masked forward pass of SPA–305.1
Decoding968.8890.2
Total968.81196.5

Sentence retrieval only takes 1.2 seconds. The overhead comes from the masked forward pass of SPA and decoding. We have added this breakdown in the revised paper(Appendix A.8).

Reviewer_31M5OpenReview ↗

Thank you for the rebuttal

Thank you for the clarifications and additional results. The fixed selection count addresses my concern, and the latency breakdown is helpful. I would like to maintain my positive rating.

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

The authors propose Agent-Radar, a training-free method for improving an LLM-based multi-agent system through context-relevant communication histories. The proposed system models agent communication as a directed graph G=(V,E,F)G=(V,E,F). Instead of pruning or compressing the communication history, Agent-Radar dynamically identifies relevant context for attention steering while retaining the full communication history available to the current agent.

A pre-trained sentence encoder, all-MiniLM-L6-v2, is used to encode the query received by the current agent and each sentence in its communication history. The proposed Agent-Radar computes the relevance score of each sentence using semantic relevance, spatial proximity, and temporal recency:

score⁡(sk∣oj(τ),vi,t,qi(t))=λsmax⁡{dG(vj,vi)−1,0}λtt−τ−1cos⁡(Enc⁡(sk),Enc⁡(qi(t))).\operatorname{score}\left(s_k\mid o_j^{(\tau)},v_i,t,q_i^{(t)}\right) = \lambda_s^{\max\{d_G(v_j,v_i)-1,0\}} \lambda_t^{t-\tau-1} \cos \left(\operatorname{Enc}(s_k),\operatorname{Enc}(q_i^{(t)})\right).

where, sks_k belongs to the output produced by agent jj at round τ\tau, dG(vj,vi)d_G(v_j,v_i) denotes the hop distance to the current agent, and t−τ−1t-\tau-1 denotes the age of the output. The parameters λs\lambda_s and λt\lambda_t control spatial and temporal decay. Sentences with scores at least θ\theta constitute the relevant context together with the current query. The relevant context is identified at each agent activation.

Following Selective Prompt Anchoring (SPA), the architecture performs a lightweight attention approximation through logit arithmetic. It contrasts the next-token logits computed from the full input with those computed from the same input with the relevant context masked. The difference is scaled in logit space to amplify attention toward the relevant context without modifying the attention weights inside the model. The main contribution is the identification of relevant context through semantic relevance, spatial proximity, and temporal recency for attention steering in multi-agent communication.

The authors use Qwen3-4B-Instruct-2507 for the main experiments, with GPTSwarm under a random communication topology as the default multi-agent system setting. The method is evaluated on HotpotQA, 2WikiMultihopQA, MuSiQue, MATH-500, and MMLU-Pro, with a reported average improvement of 4.17 absolute points over AgentDropout in the main comparison. Further experiments examine different multi-agent systems, communication topologies, and LLM backbones. Additionally, the authors perform ablation studies on MuSiQue, MATH-500, and MMLU-Pro that support the contribution of relevant-context selection and the spatial and temporal decay components and their three-run comparison reports positive mean paired differences against AgentDropout across all five benchmarks.

Critical strengths
  • The proposed Agent-Radar provides a training-free approach that identifies relevant context using a pre-trained sentence encoder together with spatial and temporal decay, and applies attention steering during inference without additional training or updates to the backbone LLM weights. Additionally, the formulation could support adaptation to custom multi-agent systems through task-specific selection of pre-trained encoders and validation of the decay factors and context-selection threshold. This potential flexibility, together with the demonstrated use of an existing pre-trained encoder, strengthens the practical value of the method.

  • The authors perform controlled ablation studies on MuSiQue, MATH-500, and MMLU-Pro that support the contribution of relevant-context selection and the spatial and temporal decay components. The decay-removal comparisons match the number of historical sentences receiving attention steering, and the query-only and full-history comparisons retain the complete transcript while using the same attention steering method. Additionally, the evaluations across GPTSwarm, AutoGen, and MAD show improvements over the corresponding multi-agent systems, while experiments with different LLM backbones and the three-run comparison against AgentDropout provide further support for the reported gains. Together, these experiments strengthen the empirical contribution by examining the individual components of Agent-Radar and its performance across different configurations.

Critical weaknesses
  • The related-work discussion would benefit from a closer comparison with automatic attention steering approaches such as AutoPASTA and Atar. AutoPASTA (Zhang et al, 2024) identifies a key sentence through LLM prompting and sentence-encoder matching, while Atar (Zhang 2026) uses the original question and current plan step to guide attention during action generation. Discussing these approaches alongside Agent-Radar would help clarify the contribution of combining semantic relevance, spatial proximity, and temporal recency in multi-agent communication. Potentially adding them to the baselines would also be useful and help strengthen the paper.
Questions for authors
  • As part of your experiments did you explore different pre-trained sentence encoders within the proposed relevance-scoring formulation? If these experiments were conducted, it would be useful to understand how encoder choice affected performance and whether the context-selection threshold or decay factors required further tuning. This would help clarify the potential for adapting Agent-Radar to task-specific multi-agent systems in a general discussion either in the paper, or here within the review.

  • Agent-Radar gives higher relevance scores to recent messages and messages from nearby agents, and Appendix C explains that useful earlier information is often restated in recent messages. Did the authors observe cases where the information needed for the answer appeared only in an earlier message or originated from a more distant agent? It would be useful to understand how Agent-Radar performed relative to semantic-only retrieval in these cases, particularly when the earlier information was not restated. This would help clarify how the proposed relevance scoring affects the use of important evidence as communication histories grow.

2 comments

Dear Reviewer oxNe,

Thank you for the positive and detailed review.

W1 Discuss and baseline against automatic attention steering methods (e.g., AutoPASTA, Atar).

Thank you for raising this related work. AutoPASTA [1] prompts the LLM to generate a key sentence for an open-book QA question, maps it back to the context with a sentence encoder, and steers attention to it. Atar [2] steers attention to the original question and the current plan step while an agent generates its next action. By contrast, Agent-Radar is designed for multi-agent communication. It combines semantic relevance with spatial and temporal decays defined over the communication topology and the discussion rounds. To compare the methods directly, we applied AutoPASTA and Atar to all agents within the same MAS.

MethodMuSiQueMATH-500MMLU-Pro
AutoPASTA32.9385.6068.80
Atar35.4686.2064.40
Agent-Radar39.7288.8069.40

Overall, Agent-Radar achieves the best performance. We have extended the related work and added this comparison in Appendix A.1.

[1] Zhang, Qingru, et al. "Model tells itself where to attend: Faithfulness meets automatic attention steering." arXiv preprint arXiv:2409.10790 (2024).

[2] Zhang, Hongxiang, Yuan Tian, and Tianyi Zhang. "Attention-aligned reasoning for large language models." Findings of EMNLP 2026.

Q1 Did you explore different pre-trained sentence encoders, and does changing the encoder require further tuning of the threshold or decay factors?

Thank you for this question. To explore the impact of the sentence encoder, we replace our default encoder with bge-base-en-v1.5:

EncoderMuSiQueMATH-500MMLU-Pro
Agent-Radar (default)39.7288.8069.40
bge-base-en-v1.538.1987.0069.77

The results show that the 'bge-base-en-v1.5' encoders perform comparably on all three tasks, showing that Agent-Radar's performance remains robust across different encoders. We have added this experiment and discussion in Appendix A.5.

Q2 How does the method perform relative to semantic-only retrieval when critical information appears exclusively in an early or distant message without being restated?

Thank you for this question. Such cases can occur, but our analysis shows that they are rare in practice. Specifically, we manually inspected 100 randomly sampled tasks. We found that only 9% of the tasks contained critical information exclusively in earlier messages or messages from more distant agents, without being restated in subsequent communication.

To understand how Agent-Radar behaves in these cases, we compare it with the original multi-agent system without steering and Agent-Radar with only semantic-based steering (no decay). We evaluated all three methods on both the 9 edge cases and the full set of 100 tasks across three runs.

Method9 edge casesAll 100 tasks
No steering3.33 ± 0.5861.67 ± 1.53
semantic-based steering6.67 ± 0.5864.00 ± 2.65
Agent-Radar6.00 ± 1.0070.33 ± 1.53

As expected, semantic-based steering performs best on the 9 edge cases (6.7 correct answers on average), as it can directly steer attention toward relevant information regardless of its temporal or spatial distance. However, it does not achieve the best overall performance across all 100 tasks, since relying solely on semantic similarity without considering recency and spatial relationship may introduce indirect noise, redundant context, or outdated information. Agent-Radar achieves the highest overall performance (70.3 correct answers on average, compared with 64.0 for semantic-based steering). Although Agent-Radar does not explicitly steer attention toward the distant critical information in the 9 edge cases, that information remains in the context window. Moreover, Agent-Radar still outperforms the original multi-agent system on the 9 edge cases (6.0 versus 3.3 correct answers on average), potentially because it explicitly steers attention toward the original question (Lines 211-212), preventing the task objective from being buried in the conversation history. We have added this analysis and discussion to Appendix C.

Thank you for providing the additional evaluations against relevant works. These comparisons thoroughly address my previous questions, and I will maintain my positive score.

Reviewer_Xy7bOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

In LLM-based MAS, multi-round interactions accumulate an increasingly long communication history, in which key context is easily diluted and hard for the current agent to use effectively, ultimately degrading system performance. To address this, the paper proposes Agent-Radar. Before each agent generates its output, Agent-Radar scores the sentences in the communication history along three dimensions: semantic relevance, the proximity of the source agent in the communication graph, and message recency. It then selects the high-priority context and uses an existing attention steering technique to amplify its influence on generation. On benchmarks spanning multi-hop question answering, mathematical reasoning, and general reasoning, Agent-Radar outperforms the strongest baseline by 4.17 points on average, and it brings gains across multiple MAS frameworks and backbone LLMs.

Critical strengths
  • The paper targets a key bottleneck that limits the scalability of MAS, namely that key information is diluted as the communication history grows, and proposes a sound direction: automatically identifying key information in the history and using attention steering to strengthen its influence on the current agent's generation. The method is training-free, incurs modest overhead, and can be plugged into different MAS. It improves performance while preserving the full communication history, thereby avoiding the information loss caused by compression or pruning.
Critical weaknesses
  • The spatial and temporal decay factors do not take the actual roles of agents into account. Multi-hop dependencies are common in practical MAS: an agent's reasoning often relies heavily on a specific, functionally related upstream role, which is not necessarily its neighbor in the communication graph. The method only considers the hop distance and age of historical messages, and ignores who sent a message and what role it plays, e.g., whether it is an error pointed out by the Critic or an intermediate result from another agent. I therefore doubt whether the method can accurately measure the importance of different messages.
  • In the paper's experiments, almost all MAS structures have each agent solve the same complete problem and then refine its answer after referring to the others' outputs. In this setting, the risk of spatial and temporal decay is low: even if a distant agent's message is down-weighted, nearby agents often provide similar content. However, if each agent is responsible for only one subtask decomposed from the query, its output is often unique and critical, and down-weighting distant messages by spatial or temporal distance could prevent key information from receiving the attention it needs. The paper does not evaluate the method on such structures.
Questions for authors
  • How does the method reflect the specific dependency needs of different roles? Could the dependencies between roles be modeled explicitly, so that the uniform decay rule is replaced by role-adaptive weighting?
  • Does the method remain effective when each agent in the MAS handles only one subtask (e.g., in task-decomposed structures such as planner–executor pipelines)?
2 comments

Dear Reviewer Xy7b,

Thank you for the detailed review.

W1/Q1 The spatial and temporal decay factors ignore agent roles. How does the method reflect specific role dependencies?

Thank you for this question. We would like to clarify that role information is already embedded into Agent-Radar through the semantic term. In our implementation, the query qi(t)q_i^{(t)} (Line 88) includes the task prompt and the current agent's system prompt, where the role is defined. Equation (7) then compares each historical sentence with this role-aware query. In such a case, the roles are reflected in the semantic weight. We have revised the paper (Line 89-91; 196-197) to clarify this point.

W2/Q2 Current experiments focus on each agent solving the same complete problem. Are there other structures being evaluated?

Thank you for this question. We would like to clarify that our evaluation in Table 2 covers diverse multi-agent architectures and orchestration patterns, including systems in which agents perform distinct, specialized roles rather than independently solving the same problem.

Specifically, GPTSwarm [1] organizes agents with different roles into a graph-based workflow. Each agent receives the original query and messages from its predecessors, while role-specific prompts define its responsibilities. For example, the Knowledge Expert identifies entities requiring further investigation, the Critic examines weaknesses in other agents' reasoning (Appendix D.4), and a separate decision node produces the final answer. Thus, agents contribute different types of information, and their outputs are not necessarily redundant.

Furthermore, AutoGen [2] supports heterogeneous agents with distinct roles and capabilities, including code generation, execution, and feedback. Unlike conventional multi-agent debate, these agents need not independently solve the same complete problem. Instead, they can perform complementary tasks and exchange intermediate results through multi-agent conversations.

As shown in Table 2, Agent-Radar consistently improves performance across all three architectures and all five benchmarks, demonstrating its effectiveness across different agent roles, communication patterns, and orchestration structures. We have updated the revised paper to include these details (Lines 296–303).

[1] Zhuge, Mingchen, et al. "Language agents as optimizable graphs." ICML 2024.

[2] Wu, Qingyun, et al. "Autogen: Enabling next-gen llm applications via multi-agent conversation." COLM 2024.

Reviewer_Xy7bOpenReview ↗

Response to the Authors

Thank you for your response. The additional clarifications have addressed most of my concerns. I've raised my current score.

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

This paper proposes AGENT-RADAR, a training-free method for improving multi-agent communication through contextaware attention steering. It combines semantic similarity, spatial proximity, and temporal recency to dynamically identify relevant communication history, and applies logit-based attention steering to emphasize the selected context. Extensive experiments across multiple benchmarks, models, and multi-agent frameworks demonstrate its effectiveness

Critical strengths

The paper addresses context dilution in multiagent communication without pruning or compressing historical information. AGENT-RADAR is training-free and can be integrated into existing multi-agent frameworks.

The experiments cover five benchmarks, multiple LLM backbones, communication topologies, and multi-agent frameworks. Multi-run comparisons and component ablations provide useful evidence of the method's effectiveness and robustness.

Critical weaknesses

The multiplicative decay and fixed relevance threshold introduce an implicit cutoff: with the default parameters, messages satisfying (a+max⁡(d−1,0)≥9)(a+\max(d-1,0)\geq9) cannot be selected for steering, regardless of semantic relevance. This may limit the method's ability to emphasize critical but distant or early information, particularly in long-horizon communication.

Table 11 shows that AGENT-RADAR increases end-to-end latency from 968.8 to 1196.5 seconds per task, approximately a 23.5% increase. The reported token counts also do not clearly separate the computational costs of sentence retrieval and additional masked forward passes required by attention steering.

Questions for authors

Can the authors provide direct evidence that the selected sentences are more relevant to successful reasoning? Can the authors provide a more detailed breakdown of retrieval, masked forward passes, and decoding costs? How does end-to-end latency scale with communication-history length and the number of agents?

2 comments

Dear Reviewer oTqA,

Thank you for the positive and detailed review.

W1: The implicit cutoff might miss critical early/distant info in long tasks.

Thank you for pointing this out. Indeed, Eq. 8 introduces a cutoff. Yet we would like to clarify that distant or early sentences above the threshold are not removed from the prompt and can still receive model attention. They are just not considered for explicit attention steering. There is always a tradeoff between coverage and precision---a lower threshold steers attention to more sentences but may introduce noise, while a higher threshold only steers attention towards the most critical sentences. This is confirmed by the experiments in Appendix A.4. The results show that lowering the threshold to steer attention to more sentences will degrade performance.

We also revisited the evaluation results to address the concern about long-horizon communication. We analyze the performance of tasks with long and short communication histories. Specifically, we compare the 150 longest communication histories (the top 25 percentile) against the 150 shortest (the bottom 25 percentile). Agent-Radar improves the F1 score over the base MAS by 4.3 points on the longest group and 4.9 points on the shortest group, so the gain on long-horizon tasks is comparable to that on short ones. We have added this analysis in the revised paper (Appendix A.6).

W2/Q2: Clarify the latency overhead breakdown

Thank you for this comment. We break down the end-to-end latency (seconds per task) into the table below:

ComponentBase MASAgent-Radar
Sentence retrieval–1.2
Masked forward pass of SPA–305.1
Decoding968.8890.2
Total968.81196.5

Sentence retrieval only takes 1.2 seconds. The overhead comes from the masked forward pass of SPA and decoding. We have added this breakdown in the revised paper(Appendix A.8).

Q1 Can the authors provide direct evidence that the selected sentences are more relevant to successful reasoning?

Thank you for this question. First, in the ablation study, we compared Agent-Rader with a variant that randomly selects sentences for attention steering. Table 5 shows that replacing our sentence selection mechanism with random selection leads to a significant performance drop. This suggests that the sentences selected by Agent-Radar contribute meaningfully to its reasoning performance.

Second, during the rebuttal period, we did a manual analysis of the relevance of the sentences selected by Agent-Rader. Specifically, we sampled 100 sentences selected by Agent-Rader and asked two annotators to independently decide whether a given sentence supports the gold answer. The two annotators found that 74.7% and 70.9% of the sentences selected by Agent-Radar were relevant to the gold answer respectively. The agreement between the two annotators is Cohen's κ\kappa = 0.86, which implies almost perfect agreement [1]. We have added this analysis to Appendix A.10.

[1] McHugh ML. Interrater reliability: the kappa statistic. Biochem Med (Zagreb). 2012;22(3):276-82. PMID: 23092060; PMCID: PMC3900052.

Q3:How does end-to-end latency scale with communication-history length and the number of agents?

Thank you for this question. We analyzed the execution logs of the scalability experiments (Lines 315-354, Figure 2) and report the corresponding end-to-end latency measurements below.

3 rounds3 agents5 agents7 agents9 agents
MAD1045180122243045
Agent-Radar1331214726053654
3 agents3 rounds5 rounds7 rounds9 rounds
MAD1045181324422858
Agent-Radar1331224030653481

As expected, end-to-end latency increases with both the number of agents and the number of communication rounds, since additional agents and rounds require more LLM calls and increase the amount of communication history to process. Importantly, Agent-Radar exhibits a scaling trend similar to that of the MAD baseline, with an average latency overhead of approximately 22% across the evaluated configurations. This suggests that Agent-Radar introduces a relatively stable overhead rather than substantially changing the scaling behavior of multi-agent communication. We have added this analysis to Appendix A.9.

Thank you for your response. The additional clarifications have addressed most of my concerns. I will keep my positive score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.