PokeAgent Challenge: Competitive Pokémon Battles as Multi-Agent Long-Context Reasoning Benchmark at Scale
arXiv 2603.15563·medium signal
arXiv 2603.15563 documents the PokeAgent Challenge (NeurIPS 2025 competition), a large-scale benchmark built on Pokémon's competitive multi-agent battle system for evaluating decision-making under uncertainty, adaptation, and long-context reasoning. The environment tests properties directly transferable to production agentic deployments: partial observability, adversarial counterplay, and long-horizon planning under combinatorial team configurations. Now fully documented with code and evaluation protocol.