acceptodds
Under review as a conference paper at ICLR 2027

Quantize the Team: Joint Quantization and Topology Adaptation for Multi-Agent LLM Systems

Abstract

Multi-agent LLM systems solve tasks by exchanging information across repeated model calls. The large number of calls and long text sequences make efficient model deployment particularly important, with post-training quantization (PTQ) serving as a key approach. However, preserving the model's performance on individual calls may not preserve the benefits of communication: a message that helped another agent before quantization may no longer improve the final answer, even though it is still delivered. We call this loss a contraction of the system's effective topology: the communication graph stays the same, but fewer connections contribute to task performance. We characterize two failure mechanisms: sending agents provide less information beyond what other agents already supply, or receiving agents make less use of useful messages. To address these failures, we propose Quantization-Aware Repair or Reroute (Q-RoR) for agents that share a low-bit model. First, we introduce a unified quantization objective that jointly preserves message generation and use. By bounding the change in each message's edge value, the drop in task score when that message is removed, it grounds their optimization in what messages contribute and avoids manually balancing separate surrogate losses. Second, we propose an activation quantization operator tailored to multi-agent collaboration: it jointly controls quantization errors across roles to preserve complementary information. We pair this operator with attention calibration to restore receivers' use of useful messages. Third, compatibility-guided rerouting goes beyond quantization on a fixed topology or topology search with a fixed model: it jointly adapts the communication topology and quantized model to restore useful communication when repair alone is insufficient. Experiments across three models and five tasks, plus two further precisions on the main model, show that Q-RoR outperforms PTQ and sequential topology-search baselines.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Quantization, LLM Agents, Multi-Agent Systems

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_dwUPOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
1
Soundness
2
Read review
Paper summary

This paper studies post-training quantization for multi-agent LLM systems. The authors measure each message’s contribution as the drop in task score when the message is removed. They diagnose two causes of the degradation under quantization: senders lose their complementary contributions, and receivers make less use of helpful messages. Based on this, they propose Q-RoR.

Critical strengths

This paper studies post-training quantization for multi-agent LLM systems. The authors measure each message’s contribution as the drop in task score when the message is removed. They diagnose two causes of the degradation under quantization: senders lose their complementary contributions, and receivers make less use of helpful messages. Based on this, they propose Q-RoR.

Critical weaknesses

The paper frequently switches between different terms and referents for the same concepts. This substantially reduces readability, and understanding the method required considerable extra effort. This work frames its goal as the joint optimization of graph structure and model quantization. However, the experiments use only a single six-agent workflow, and the method is not evaluated under broader agent/graph settings. In addition, the optimized graph structure is reported only for Qwen3-8B (G1 in Appendix C.6); the optimized graphs for the other settings are not given. Taken together, these two points make the final results look more like an engineering advantage, and they do not fully match the motivation and findings developed at length in the earlier sections. The paper reports only inference runtime (Q5, Table 6), and it does not include the inference time of the FP model as a reference.

Questions for authors
  1. Could the authors report the optimized graph structures for all settings, and analyze why these optimized structures may improve multi-agent performance under quantization?
  2. Could the authors report the inference time of the FP model as a reference in Table 6?
  3. When discussing algorithmic efficiency, would it be more appropriate to compare the offline calibration and graph-search time?
Reviewer_VKc8OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies how post-training quantization changes the usefulness of communication in multi-agent systems. Q-RoR calibrates a shared quantized model using transcript KL on original and edge-ablated workflows, repairs sender activation quantization and receiver attention, and reroutes messages with candidate-specific recalibration. Experiments evaluate three models on five tasks at W4A8, with additional W8A8 and W4A4 results for Qwen3-8B.

Critical strengths

This paper studies the efficient multi-agent deployment problem: preserving the contribution of inter-agent communication after model quantization. It connects quantization and topology adaptation through task-level edge values, providing a basis for diagnosing and repairing communication losses.

Critical weaknesses
  1. Insufficient Motivation Claim. The introduction motivates PTQ through the large token consumption of multi-agent systems, but does not clearly distinguish reducing communication volume from reducing the resource cost of model execution. Existing methods could tackle the problem of efficient multi-agent deployment by pruning agents or communication edges, as in AgentPrune [1] and AgentDropout [2], or by enabling latent communication, as in LatentMAS [3] and KVComm [4].

[1] Cut the Crap: An Economical Communication Pipeline for LLM-based Multi-Agent Systems. ICLR, 2025.
[2] AgentDropout: Dynamic Agent Elimination for Token-Efficient and High-Performance LLM-Based Multi-Agent Collaboration. ACL, 2025.
[3] Latent Collaboration in Multi-Agent Systems. ICML, 2026.
[4] KVComm: Enabling Efficient LLM Communication through Selective KV Sharing. ICLR, 2026.

  1. Unclear Positioning Against Prior Work. The Introduction characterizes existing approaches to post-training quantization and multi-agent topology optimization without citing representative studies. In particular, the discussion of their limitations is not clearly supported by references to specific methods, which affects the rigor of the paper's expression and the credibility of innovative claims.

  2. Insufficient Experiments. Although reduced multi-agent inference cost is a central application motivation, the evaluation primarily establishes task-performance preservation under quantization. Missing its actual deployment savings over FP, such as end-to-end latency, model storage and peak GPU memory on the same hardware and matched workloads.

  3. Unsupported Code-Availability Statement. The reproducibility statement explicitly states that code is available through an anonymous repository. However, the repository snapshot available for inspection contains only a short README identifying the method, without any code.

Questions for authors

N/A

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

This paper proposes Q-RoR, a communication-aware quantization framework that preserves the utility of inter-agent messages through edge-value-aware calibration, role-specific quantization repair, and topology rerouting for low-bit multi-agent LLM deployment.

Critical strengths

The paper identifies an interesting problem. I like the idea that preserving per-call model fidelity does not necessarily preserve the marginal utility of inter-agent communication.

Critical weaknesses

codes are not open. empty link, fake access.

Questions for authors

None

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.