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

Shallow read  ·  2026  ·  source  ·  all reading

L-006 L-010

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

Source: cs.MA updates on arXiv.org — https://arxiv.org/abs/2607.03473 Date read: 2026-09-01 Connected to: L-006, L-010, seed-048 Kind: content Escalation: store-only Escalation rationale:

What this is

A multi-agent reinforcement learning paper proposing a method ("MUTE") that optimizes communication in bandwidth-constrained cooperative games by filtering messages based on task return rather than information-theoretic metrics. The core claim is that high-informativeness and task-relevance are decoupled — a message can be maximally informative yet irrelevant to joint reward.

What I took from it

The paper confirms a key friction in L-006 (Coordination Cost Conservation): when bandwidth becomes a hard constraint, the system does not eliminate coordination cost, it displaces it. Here, information-theoretic surrogates (mutual information, entropy reduction) are abandoned in favor of return-preserving filtering. This is not a solution to coordination cost; it is a relocation — computational overhead shifts from communication selection to return prediction and counterfactual evaluation.

The mechanism described aligns with L-004 (Goodhart Generalization): optimizing for "informativeness" as a proxy for "useful coordination" produces messages that are high-signal in the statistical sense but functionally inert. The paper's intervention — switching the proxy to "return-relevant" — is pragmatically correct but does not resolve the underlying tension: in partially observable settings, what constitutes "return relevance" is itself only observable post-hoc, creating a moving target under online adaptation.

The work sits at the edges of L-010 (Coordination Adoption Nonmonotonicity) but does not substantively engage it. The sparse communication regime could produce adoption cascades or collapse, depending on whether agents can distinguish signal from filtered noise — the paper does not address this.

Research connections

  • L-006: Confirms coordination cost displacement rather than elimination; bandwidth constraint moves burden to return-prediction layer rather than removal.
  • L-004: Documents Goodhart failure in communication sparsity (informativeness ≠ return relevance); proxy switching is pragmatic but does not resolve the decoupling.
  • L-010: Touches the boundary (sparse signals + conditional adoption) but does not model nonmonotonic adoption dynamics.
  • seed-048: Weak connection — paper optimizes within communication protocol, does not examine how capability and cooperation invert under bandwidth.

Seed

Seed title: Proxy Decoupling Under Legibility Collapse Seed type: observation Seed text: In resource-constrained coordination systems, proxies that are optimal under full observability (information-theoretic metrics) become misaligned with task objectives as constraints tighten. Switching to a "truer" proxy (return-relevance) does not resolve the decoupling — it relocates the cost and creates a nested optimization problem (predicting what is return-relevant without full state visibility). Under progressive constraint tightening, the cost of maintaining proxy alignment may eventually exceed the cost of the constraint itself, creating a regime transition. This pattern likely generalizes beyond communication to any coordination layer where legibility decays faster than the underlying system dynamics.