TextArena 0.7.0

A Collection of Competitive Text-Based Games for Language Model Evaluation and Reinforcement Learning

A suite of 80+ Single-/Two-/Multi-Player texted based games for benchmarking and training of LLMs.

Play | Leaderboard | Games | Examples

Updates

• 14/07/2025 Announcing MindGames a NeurIPS2025 competition for training LLMs on various TextArena games that require theory of mind.
• 01/07/2025 Release of v0.6.9 with 100 games and simplified states, new observation wrappers for training and default wrappers for environments.
• 01/07/2025 Release of SPIRAL: Self-Play on Zero-Sum Games Incentivizes Reasoning via Multi-Agent Multi-Turn Reinforcement Learning introducing RL via self-play on TextArena games as a potential new training paradigm.
• 22/06/2025 Release of UnstableBaselines a light weight async online RL library for training LLMs on TextArena games.
• 16/04/2025 Release of the TextArena paper
• 14/02/2025 Release of the new, stable version for both pip and the website
• 31/01/2025 Initial demo release highlighted by Andrej Karpathy (crashing all our servers)

Introduction

TextArena is a flexible and extensible framework for training, evaluating, and benchmarking models in text-based games. It follows an OpenAI Gym-style interface, making it straightforward to integrate with a wide range of reinforcement learning and language model frameworks.

Getting Started

Installation

Install TextArena directly from PyPI:

pip install textarena

Offline Play

The only requirement Agents need to fulfill is having a call function that accepts string observations and returns string action. We have implemented a number of basic agents that you can find [here](TODO link). In this example, we show how you can let GPT-4o-mini play against anthropic/claude-3.5-haiku in a game of TicTacToe.

We will be using the OpenRouterAgent, so first you need to set you OpenRouter API key:

export OPENROUTER_API_KEY="YOUR_OPENROUTER_API_KEY"

Now we can build the models and let them play:

import textarena as ta

# Initialize agents
agents = {
    0: ta.agents.OpenRouterAgent(model_name="GPT-4o-mini"),
    1: ta.agents.OpenRouterAgent(model_name="anthropic/claude-3.5-haiku"),
}

# Initialize the environment
env = ta.make(env_id="TicTacToe-v0")

# wrap it for additional visualizations
env = ta.wrappers.SimpleRenderWrapper(env=env) 

env.reset(num_players=len(agents))

done = False
while not done:
    player_id, observation = env.get_observation()
    action = agents[player_id](observation)
    done, step_info = env.step(action=action)

rewards, game_info = env.close()

Citation

If you use TextArena in your research, please cite:

@misc{guertler2025textarena,
    title={TextArena}, 
    author={Leon Guertler and Bobby Cheng and Simon Yu and Bo Liu and Leshem Choshen and Cheston Tan},
    year={2025},
    eprint={2504.11442},
    archivePrefix={arXiv},
    primaryClass={cs.CL},
    url={https://arxiv.org/abs/2504.11442}, 
}

How to Contribute:

If you have any questions at all, feel free to reach out on discord. The below issues are great starting points if you want to contribute:

• Transfer the 'How to Contribute' from here to individual issues
• Make RushHour board generation algorithmic
• extend 2048 to arbitrary board sizes (should be very straight forward)
• extend Fifteenpuzzel to arbitrary sizes
• Add a nice end-of-game screen to the SimpleRenderWrapper visualizations