The PokéAgent Challenge, presented as a NeurIPS 2025 competition, establishes a novel research benchmark centered on the Pokémon video game series to advance AI decision-making. The project seeks to synthesize research across reinforcement learning (RL), large language models (LLMs), game theory, and planning by utilizing Pokémon as a testbed for agents navigating complex, partially observable, and long-horizon environments.

Core Methodology and Research Objectives

The competition is bifurcated into two primary tracks, each addressing distinct challenges in contemporary AI:

1. Competitive Battling (Track 1):
This track focuses on the intricacies of competitive game theory and agent-environment interaction under partial observability. The methodology necessitates the development of agents capable of robust opponent modeling. Because the opponent's hidden state (e.g., specific Pokémon move-sets, held items, or remaining team composition) is not directly accessible, agents must employ probabilistic reasoning to infer the opponent's strategy. Mathematically, this can be framed as a Partially Observable Stochastic Game (POSG). Participants are encouraged to develop models that predict hidden latent variables $z_t$ based on the history of observations $o_{1:t}$ and past actions $a_{1:t-1}$:
$$P(z_t | o_{1:t}, a_{1:t-1})$$
Successful agents must adaptively update these beliefs to optimize expected utility in a high-stakes, adversarial environment.

2. RPG Speedrunning (Track 2):
This track targets the challenge of long-horizon planning. In this setting, agents must navigate massive state spaces over thousands of timesteps to achieve specific objectives efficiently. The primary technical hurdle involves maintaining coherent, multi-stage planning where early decisions drastically affect outcomes in the distant future. This requires architectural innovations in temporal credit assignment and memory management, moving beyond simple reactive policies toward hierarchical planning architectures. Agents must maximize a cumulative reward function $R = \sum_{t=0}^{T} \gamma^t r_t$, where $T$ represents an exceptionally large horizon, forcing the integration of external knowledge bases and long-context reasoning.

Technical Contributions and Standardization

The PokéAgent Challenge aims to create a unified research framework by providing standardized datasets and starter code. By utilizing Pokémon, the authors propose a medium that bridges the gap between simulated environments and real-world complexity, allowing for the evaluation of:

• Strategic Adaptation: The capability of an agent to generalize across varied configurations and unexpected tactical shifts by the opponent.
• Knowledge Integration: Methods for augmenting RL policies or LLM decision-making with structured game data and external reference materials.
• Robustness: Assessing how models perform when subjected to non-stationary environments where optimal strategies evolve based on meta-gaming.

The initiative highlights the necessity of combining diverse paradigms—specifically, symbolic reasoning, neural-based RL, and LLM-driven inference—to solve tasks that demand both strategic depth and sequential coherence. Winning entries will be presented at the NeurIPS 2025 workshop, with the ultimate goal of producing a comprehensive, multi-authored report on the state of autonomous agency in complex decision-making tasks.