Meta-Knowledge Roles: Evolving Specialization for Knowing-When-to-Know

Eisenstein et al. show that multi-agent reward can induce meta-knowledge. Building on role-centric MARL (SMR; Jin & Liu, 2024) and heterogeneous embodied collaboration (Liu et al., 2023), this idea learns roles that encode “who is best at abstaining,” “who is best at tool verification,” and “who is best at parametric recall.” Roles are inferred from interaction trajectories via mutual information objectives (as in SMR), but the objective targets meta-behaviors rather than task primitives. This differs from existing multi-agent role work by focusing on “epistemic roles” aligned with selective prediction and tool-use calibration. The promise is that such roles will both improve team performance and produce interpretable templates for downstream single-agent deployment: we can distill a solo agent with a modular “abstainer head,” “verifier head,” and “retriever head,” improving transfer across domains like MACT for TQA (Zhou et al., 2024) and GraphTeam (Li 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. Multi-Agent Self-Motivated Learning via Role Representation. Yuchao Jin, Quan Liu (2024). IEEE International Joint Conference on Neural Network.
3. Heterogeneous Embodied Multi-Agent Collaboration. Xinzhu Liu, Di Guo, Xinyu Zhang, Huaping Liu (2023). IEEE Robotics and Automation Letters.
4. Efficient Multi-Agent Collaboration with Tool Use for Online Planning in Complex Table Question Answering. Wei Zhou, Mohsen Mesgar, Annemarie Friedrich, Heike Adel (2024). North American Chapter of the Association for Computational Linguistics.
5. GraphTeam: Facilitating Large Language Model-based Graph Analysis via Multi-Agent Collaboration. Xin Li, Qizhi Chu, Yubin Chen, Yang Liu, Yaoqi Liu, Zekai Yu, Weize Chen, Cheng Qian, Chuan Shi, Cheng Yang (2024). arXiv.org.