Artificial Pokemon Intelligence in the PokeAgent Challenge (pokeagent.github.io)

NeurIPS 2025 organizers have launched the PokeAgent Challenge, a competition that uses Pokémon as a controlled, rich testbed to push AI decision-making. The challenge offers two complementary tracks—competitive battling (multiplayer, partial observability, opponent modeling) and RPG speedrunning (long-horizon planning and sequential decision-making)—with starter code, datasets, submission guidelines, and a community Discord for team coordination. Compute credits have already been awarded to approved teams; multiple prize categories reward leaderboard performance, bracket winners, and novel approaches (e.g., Best LLM-based method, Best RL method). Winning teams may receive workshop invitations at NeurIPS and co-authorship on the competition report. Technically the PokeAgent Challenge is significant because it provides a standardized benchmark bridging reinforcement learning, game theory, planning, and large language models. By combining adversarial, multi-agent partial-observability scenarios with long-context single-agent speedruns, the contest stresses opponent modeling, memory and long-horizon reasoning, hierarchical planning, and model-based vs. model-free tradeoffs—capabilities that many current benchmarks neglect. The competition’s datasets and baseline code enable reproducible evaluation and method comparison, while method-specific prizes and sponsor outreach aim to accelerate cross-paradigm innovation and real-world applicability of AI systems.