Cross-Domain Meta-Knowledge Transfer: From Self-Play to Drug Discovery, Graphs, and Tables

Eisenstein et al. demonstrate transfer to solo tool use. We extend this by creating MetaBench: multi-domain tasks that plug a learned “selective prediction and tool triage” module into DrugAgent (Liu et al., 2024), GraphTeam (Li et al., 2024), and MACT for table QA (Zhou et al., 2024). The key idea is to pretrain meta-knowledge policies in a synthetic self-play sandbox with heterogeneous tools (e.g., noisy retrievals, code exec, domain KBs) and then reuse them as domain-agnostic controllers. This differs from task-specific multi-agent pipelines by decoupling meta-knowledge learning from domain skill learning. We hypothesize improved robustness to tool failures and better budgeted accuracy across domains, offering a principled way to avoid domain overfitting and accelerate deployment.

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. DrugAgent: Automating AI-aided Drug Discovery Programming through LLM Multi-Agent Collaboration. Sizhe Liu, Yizhou Lu, Siyu Chen, Xiyang Hu, Jieyu Zhao, Tianfan Fu, Yue Zhao (2024). arXiv.org.
3. 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.
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.