.png)
3 weeks ago
2
A NeurIPS 2025 competition advancing AI decision-making through Pokémon. Featuring competitive battling and RPG speedrunning tracks to unify research in reinforcement learning and large language models.
⏰ Competition Ends In
-- Days
-- Hours
-- Minutes
-- Seconds
Final submissions close November 15th, 2025
Competition Tracks
Click to explore detailed information, submission guidelines, starter code, and datasets for each track.
About the Competition
Scientific Relevance
The PokéAgent Challenge positions Pokémon as an ideal testbed for artificial intelligence research, offering two complementary tracks that address fundamental challenges in decision-making.
This competition addresses critical frontiers in AI research at the intersection of reinforcement learning, game theory, planning, and language models. It creates a standardized benchmark for opponent modeling under partial observability and long-horizon reasoning—two capabilities essential for advancing AI beyond controlled environments.
Key Research Areas
- Opponent Modeling: Track 1 requires sophisticated opponent modeling under partial observability.
- Long-Horizon Planning: Track 2 challenges agents to maintain coherent planning across thousands of timesteps.
- Strategic Adaptation: Both tracks require agents to generalize across varied scenarios and adapt to novel situations.
- Knowledge Integration: Opportunity to develop methods that augment decision-making with existing reference materials.
Ready to Get Started?
Join the PokéAgent Challenge Discord server to register and connect with other participants!
Compute Credit Application: Application closed and credits awarded. All compute credits have been distributed to approved teams.
Prizes & Recognition
Multiple prize categories ensure diverse contributions are recognized and rewarded
Highest Ranking Teams
Performance-based prizes for teams placing high on the leaderboard in each track. Best agents in Track 1 tournament brackets and Track 2 speedrun rankings will receive prizes and recognition.
Research Impact Awards
Method-specific prizes recognizing innovative approaches across all AI paradigms. Categories may include: Best LLM-based method, Best RL method, or other breakthrough techniques.
NeurIPS 2025 Presentation
Workshop invitations for winning teams to be highlighted at NeurIPS 2025. Selected teams will also be invited as co-authors on the official competition report publication.
$15,000+ Total Prize Pool
Distributed across multiple categories to reward both excellence and innovation
Organizing Team
Seth Karten
Princeton University
Jake Grigsby
UT Austin
Stephanie Milani
Carnegie Mellon University
Kiran Vodrahalli
Google DeepMind
Amy Zhang
UT Austin
Fei Fang
Carnegie Mellon University
Yuke Zhu
UT Austin
Chi Jin
Princeton University
Cite This Work
If you use this competition in your research, please cite our paper:
