Communication-Budgeted Self-Play for Minimalist Meta-Knowledge

Eisenstein et al.’s societies collaborate freely; in practice, communication is constrained. Drawing on network-based message selection (Hyeon et al., 2024), we propose a self-play regime that penalizes message count, bandwidth, and latency, forcing teams to compress epistemic state and uncertainty signals. The novelty lies in optimizing the content and timing of meta-knowledge messages (e.g., “I’m uncertain; request tool X”) under tight budgets, yielding communication-efficient strategies that still improve solo selective prediction. This conflicts productively with the assumption that more debate is always better, and it provides a pathway to deploy agent teams in edge or real-time settings. A byproduct is a catalog of learned communication primitives (confidence pings, trust flags) that can be plugged into frameworks like LLM-Collab (Cao et al., 2024).

References:

1. Don't lie to your friends: Learning what you know from collaborative self-play. Jacob Eisenstein, Reza Aghajani, Adam Fisch, Dheeru Dua, Fantine Huot, Mirella Lapata, Vicky Zayats, Jonathan Berant (2025). arXiv.org.
2. Network-Based Message Selection for Efficient Communication of Multi-Agent Collaboration. Seungjun Hyeon, Minho Park, Dong-oh Kang (2024). Information and Communication Technology Convergence.
3. LLM-Collab: a framework for enhancing task planning via chain-of-thought and multi-agent collaboration. Hong Cao, Rong Ma, Yanlong Zhai, Jun Shen (2024). Applied Computing and Intelligence.