MUTE Return-Preserving Communication Unlearning for Efficient Multi-Agent Coordination

Jul 16, 2026

Upload by Ivy

Inter-agent communication is critical for coordinating Multi-Agent Reinforcement Learning agents under partial observability, yet real deployments face strict bandwidth limits that favor sparse interactions. Existing methods often optimize information-theoretic surrogates, but they can select messages that are statistically informative while being irrelevant to the joint task return. MUTE formulates communication reduction as value-guided machine unlearning, estimates Counterfactual Message Value via attention, and unlearns low-value message transmissions while preserving the original joint policy’s return. Experiments show 80%–90% bandwidth reduction with comparable performance and a theoretical bound on return degradation.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages24

Reading Time1 h 0 min

Introduction

Background: CTDE and communication under partial observability

Bandwidth limits and information-theoretic reduction methods

Proposed approach: MUTE and counterfactual message value

Unlearning objective and return-preserving guarantees

MUTE: Return-Preserving Communication Unlearning for Efficient Multi-Agent Coordination Rui Zuo Syracuse University [rzuo02@syr.edu](rzuo02@syr.edu) Qinwei Huang Syracuse University [qhuang18@syr.edu](qhuang18@syr.edu) Mingyang Li Syracuse University [mli170@syr.edu](mli170@syr.edu) Zhenhang Zhang Syracuse University [zzhan281@syr.edu](zzhan281@syr.edu) arXiv :2607 .03473v 1 [ cs .MA] 3 Jul 2026 Simon Khan Air Force Research Laboratory [simon.khan@us.af.mil](simon.khan@us.af.mil) Qinru Qiu Syracuse University [qiqiu@syr.com](qiqiu@syr.com) Abstract Inter-agent communication is critical for coordinating Multi-Agent Reinforcement Learning (MARL) agents under partial observability to perform effectively in cooperative games; however, real-world bandwidth constraints demand sparse interactions. Prior approaches primarily address this trade-off by optimizing informationtheoretic surrogates. We argue that these statistical proxies are fundamentally misaligned with the true objective: a message can be highly informative yet irrelevant to the joint return of the task. In this work, we propose Message Unlearning for Targeted Efficiency (MUTE), a framework that views communication reduction as a value-guided machine unlearning problem. MUTE rigorously quantifies the Counterfactual Message Value using an attention-based estimator, and systematically unlearns the transmission of low-value messages from a policy trained without any communication constraints. This is achieved through a dual-objective mechanism that enforces communication sparsity while preserving the return of the original joint policy. We derive a theoretical upper bound on the performance gap induced by this sparsification, guaranteeing controlled return degradation. We also empirically evaluate MUTE on various complex multi-agent environments, achieving 80% to 90% bandwidth reduction while maintaining performance comparable to state-of-the-art baselines. 1 Introduction Multi-Agent Reinforcement Learning (MARL) has emerged as a powerful paradigm for solving complex coordination problems, ranging from autonomous vehicles [1] to micromanagement in StarCraft II [2] . To address the challenges of non-stationarity and scalability, the Centralized Training with Decentralized Execution (CTDE) [3] framework has become the standard, exemplified by valuebased methods [4, 5] and actor-critic approaches [6–8] . However, in environments characterized by partial observability and policy-induced non-stationarity, decentralized execution often falters. Communication emerges as a pivotal solution to this challenge [8, 9], empowering agents to transcend the limitations of local perception by exchanging diverse signals—ranging from raw observations [10, 11] and latent intentions [5] to encoded experiences [12] . This continuous information exchange serves to stabilize the learning process against environmental non-stationarity, granting agents a clearer understanding of the global state [5, 13] . While unrestricted communication can mitigate partial observability, bandwidth remains a scarce resource in real-world deployments [14, 15] . Prior works have primarily addressed this constraint by optimizing information-theoretic [14, 13, 5] auxiliary objectives to reduce communication. For Preprint. instance, IMAC [14] enforces communication sparsity via the Information Bottleneck principle, explicitly minimizing the mutual information between observations and messages, which implicitly assumes that high-entropy messages are redundant. NDQ [13] and MAIC [5] prioritize messages by maximizing the mutual information between the message and the recipient’s action selection. This promotes messages that alter behavior, regardless of whether that alteration leads to higher returns. By optimizing for statistical significance without directly targeting the joint return, these methods risk prioritizing messages that are merely informative over those that are truly

FAQs

Why do information-theoretic communication reduction methods underperform for return optimization?

They prioritize messages that are statistically informative, which may not translate into increased joint return. A message can be highly informative yet irrelevant to the task’s joint outcome.

What is MUTE and how does it reduce communication?

MUTE treats communication reduction as a value-guided unlearning problem. It estimates Counterfactual Message Value to identify redundant low-value messages and then unlearns their transmission while preserving the joint return.

How does MUTE estimate the contribution of a message without exponential cost?

MUTE uses a Message-Value Estimator with an attention-based approach to approximate the Counterfactual Message Value, comparing message-present versus message-masked counterfactual scenarios.

Jul 16, 2026

Upload by Ivy