Reproduce “COvolve: Adversarial Co-Evolution of Large-Language-Model-Generated Policies and Environments via Two-Player Zero-Sum Game” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Alkis Sygkounas, Rishi Hazra, Andreas Persson, Pedro Zuidberg Dos Martires, Amy Loutfi

COvolve introduces a framework for adversarial co-evolution, employing Large Language Models to generate both training environments and agent policies within a two-player zero-sum game. This approach produces progressively complex environments and robust, generalizable policies across diverse domains, effectively mitigating catastrophic forgetting through Mixed-Strategy Nash Equilibrium computation.

Key insights

  • Adversarial co-evolution, driven by Large Language Models, can automatically induce progressively complex curricula for training AI agents.
  • The use of Mixed-Strategy Nash Equilibrium (MSNE) is crucial for building robust meta-policies that effectively prevent catastrophic forgetting across a growing archive of tasks.
  • Large Language Models are capable of synthesizing complex executable code for both environments and agent policies, offering advantages in generalizability and interpretability over traditional neural encodings.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.