Dataset Card for Connect4 Reasoning Task

1. Dataset Summary

This dataset is dynamically constructed using the GAMEBoT framework in the Connect4 game. It is designed to evaluate Large Language Models' (LLMs) ability in symbolic reasoning, board state parsing, and lookahead planning.

By serializing 6 × 7 grid states into text-based formats, this dataset challenges models to identify winning topologies in a deterministic environment with a state-space complexity of approximately 4.5 × 10¹², ensuring a robust evaluation against data contamination.

2. Construction Methodology

• State Generation: Two randomized agents simulated gameplay to generate a diverse distribution of board states, covering both balanced and critical tactical scenarios.
• Filtering and Balancing:
  • Duplicate states were removed.
  • To prevent label imbalance (where LLMs might score high by simply predicting "no winning moves"), states with empty answers were downsampled to a maximum ratio of 20%.
• Ground Truth: All labels were verified using a perfect solver integrated within the GAMEBoT engine.

3. Evaluation Protocol

For each instance, the model is prompted to solve two sub-problems:

1. Self-Winning: "Are there any potential winning moves to form 4-in-a-row for you? Output all winning moves."
2. Opponent-Winning: "Are there any potential winning moves to form 4-in-a-row for your opponent? Output all winning moves."

The evaluation extracts the final answer from the model's Chain-of-Thought (CoT) and compares it against the engine-validated ground truth.

4. Citation

If you find this dataset helpful in your research, please cite the following work:

@inproceedings{lin2025gamebot,
  title={GAMEBoT: Transparent assessment of LLM reasoning in games},
  author={Lin, Wenye and Roberts, Jonathan and Yang, Yunhan and Albanie, Samuel and Lu, Zongqing and Han, Kai},
  booktitle=ACL,
  year={2025}
}