Chess

Description

Chess is an environment for evaluating agents on playing chess against Stockfish. Agents play full games of chess by submitting moves in UCI notation. Stockfish responds at a configurable skill level (1-20). The environment provides two sub-environments: ChessTextEnv (FEN text observations) and ChessImageEnv (rendered board image observations). Reward is computed per move using a logistic mapping of Stockfish's centipawn evaluation.

Capabilities

• Playing chess against an engine at varying difficulty levels
• Strategic planning and tactical reasoning in chess
• Understanding FEN notation and UCI move format
• Multi-turn decision-making in a competitive game setting

Compute Requirements

Chess requires 4 GB RAM and 4 CPUs to run the Stockfish chess engine efficiently. The Stockfish binary must be available on the server.

License

GPL-3.0 (due to Stockfish engine dependency).

Tasks

There is one split: train (40 tasks). Tasks are parameterized by two dimensions:

• Skill level (1-20): Controls Stockfish's playing strength.
• Player color (white or black): Determines which side the agent plays.

This gives 20 skill levels x 2 colors = 40 tasks.

Reward Structure

This is a dense reward environment with continuous scoring. After each move, the environment evaluates the board position using Stockfish (depth 8) and maps the centipawn score to a reward in [-1, 1] using a logistic function:

$\text{reward} = 2 \cdot \sigma(k \cdot \text{cp}) - 1$

where $k = 0.004$ and cp is the centipawn evaluation from the agent's perspective. Mate detection maps to +/-1.0. Invalid moves receive a reward of -1.0.

We do not use LLM graders for this task.

Data

No external data is required. Games are played in real time against the Stockfish engine.

Tools

Agents are given a single tool across both sub-environments:

• step: Submit a move in UCI format (e.g., "e2e4"). Returns the updated board state (FEN text in ChessTextEnv, board image in ChessImageEnv) after Stockfish responds. The game ends when a checkmate, stalemate, or draw condition is reached.

Time Horizon

Chess is a multi-turn environment. Each task is a full game of chess, with the agent and Stockfish alternating moves until the game ends.

[How many average tool calls?]

Environment Difficulty

[Statistics on environment difficulty here]

Other Environment Requirements

There are no further environment requirements; Chess works out of the box with the OpenReward endpoint.

Safety

Agents in Chess play chess against a Stockfish engine. The environment does not present direct safety risks, as agents only submit chess moves with no access to external systems.

Citations

@dataset{GRChess,
  author    = {General Reasoning Inc. Team},
  title     = {Chess},
  year      = {2026},
  publisher = {OpenReward},
  url       = {https://openreward.ai/GeneralReasoning/Chess}
}