Evolutionary Populations for Epistemic Diversity and Generalization

Beyond single-society training (Eisenstein et al.), we borrow from early evolutionary multi-agent work (Cetnarowicz et al., 1996) and population-based self-play (Huynh et al., 2024) to maintain a heterogeneous ecosystem of societies. Selection favors teams that generalize across tool perturbations, domain shifts, and budget regimes. We hypothesize that epistemic diversity in training partners yields solo agents with better out-of-distribution selective prediction and tool trust calibration. This departs from monolithic pipelines like LLM-Collab or MACRec (Wang et al., 2024) by explicitly preserving strategic diversity rather than converging on a single best policy. The research would quantify how diversity in meta-knowledge strategies translates into robustness and sample efficiency.

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. The Application of Evolution Process in Multi-Agent World to the Prediction System. K. Cetnarowicz, Marek Kisiel-Dorohinicki, E. Nawarecki (1996).
3. Multi-Agent Training for Pommerman: Curriculum Learning and Population-based Self-Play Approach. Nhat-Minh Huynh, Hoang-Giang Cao, I-Chen Wu (2024). arXiv.org.
4. MACRec: A Multi-Agent Collaboration Framework for Recommendation. Zhefan Wang, Yuanqing Yu, Wendi Zheng, Weizhi Ma, Min Zhang (2024). Annual International ACM SIGIR Conference on Research and Development in Information Retrieval.
5. 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.