MUTE Teaches Multi-Agent Systems to Unlearn Wasteful Communication

Source: arXiv — 2026-07-06

Summary

Rui Zuo and coauthors propose MUTE, a method for selectively "unlearning" inefficient or redundant communication patterns in multi-agent reinforcement learning while provably preserving task return. Rather than the usual compression or pruning of messages, MUTE targets specific communication behaviors for removal and gives a formal guarantee that performance doesn't degrade as a result.

Key Takeaways

• Targets specific redundant/wasteful communication patterns for removal rather than uniformly compressing all messages.
• Provides a formal return-preservation guarantee — unlearning communication shouldn't come at the cost of task performance.
• A fresh angle on multi-agent communication efficiency compared to typical bandwidth-compression or message-pruning approaches.
• Relevant as multi-agent RL systems scale up and communication overhead becomes a real bottleneck, not just a theoretical concern.

Discussion

(No questions yet — ask follow-ups via a Claude Code chat session on this repo; answers get appended here.)