Resource-constrained Amazons chess decision framework integrating large language models and graph attention

Tianhao Qian, Zhuoxuan Li, Jinde Cao, Xinli Shi, Hanjie Liu, Leszek RutkowskiarXiv 2026 · DOI

TL;DR

This paper introduces a novel, lightweight framework for the game of Amazons that leverages Large Language Models (LLMs) for data generation and Graph Attention Networks (GATs) for structural reasoning. The approach achieves competitive performance in resource-constrained environments, outperforming its teacher model and individual components by effectively filtering noisy LLM outputs and learning robust decision strategies.

What Problem Does This Paper Address?

Developing competitive AI agents for complex games like Amazons typically requires vast amounts of expert data and significant computational resources. This poses a challenge for deployment on everyday devices or in environments with limited computational budgets. Existing deep learning methods often struggle with data scarcity and high training costs. Furthermore, while LLMs can generate synthetic data, this data can be noisy and imperfect, containing hallucinations or inconsistencies. The paper aims to address this by creating a resource-efficient framework that can learn strong decision-making capabilities from imperfect LLM-generated supervision, without relying on extensive expert demonstrations.

Methods

The proposed framework is a hybrid approach that integrates several key components:

  • Monte Carlo Tree Search (MCTS) with an Update Mechanism: MCTS forms the search backbone. A novel "global depth normalization" mechanism is introduced to rescale node values based on their depth, making nodes from different layers more comparable and mitigating the impact of estimation noise in deep branches.
  • UCT-AE (Upper Confidence Bound - Autoencoder): This module refines handcrafted evaluation features within the MCTS search. It uses autoencoders to learn a compressed, low-dimensional representation of game state features, improving evaluation efficiency without significant latency. Two distinct autoencoders are used for movement and placement evaluation.
  • GAT-AE (Graph Attention Network - Autoencoder): This component exploits the structural information within the MCTS search tree. It transforms the search tree into a graph representation and uses Graph Attention Networks to capture relational dependencies among candidate moves, enhancing discrimination.
  • Stochastic Graph Genetic Algorithm (SGGA): SGGA is employed for stochastic candidate selection on the graph derived from the MCTS tree. It uses a genetic procedure to explore the graph, introducing diversity and robustness to the candidate selection process.
  • LLM-based Data Generation: GPT-40-mini is used to generate synthetic supervision signals (scores for candidate moves). Crucially, these scores are not used directly as labels but are transformed into probabilistic selection signals by SGGA, making the training process more robust to noise.

Datasets

The paper does not explicitly mention or introduce new datasets. Instead, it relies on synthetic data generated by the GPT-40-mini LLM for training. The evaluation is performed on the Game of the Amazons itself, with experiments conducted as head-to-head games.

Key Insights and Ideas

  • Weak-to-Strong Generalization in Games: The research demonstrates that a general-purpose LLM can serve as a "weak supervisor" to train a specialized, competitive AI agent for a complex game, even when the LLM's output is noisy.
  • Structural Bias for Robustness: By transforming LLM outputs and search states into graph structures, the framework imposes an inductive bias that encourages learning stable structural patterns (connectivity, territorial interaction) rather than memorizing noisy labels. This acts as a form of implicit denoising.
  • Complementarity of Components: The success of the framework is attributed to the synergistic interaction of its components. Lightweight evaluation refinement (UCT-AE), stochastic exploration (SGGA), and structural modeling (GAT-AE) work together, with no single module being solely responsible for the performance gains.
  • Resource Efficiency is Achievable: The framework demonstrates that competitive performance in complex games can be achieved with significantly fewer resources than traditional approaches, making AI more accessible.

Key Results

  • The proposed hybrid framework consistently outperforms its individual components (UCTS-AE, SGGA, GAT-AE) in ablation studies.
  • Against the teacher model GPT-40-mini, the framework achieves a competitive win rate of 66.5% with a modest increase in search budget (N=50), compared to 45.0% at N=30.
  • The framework achieves win rates between 57.5% and 79.5% against ablation baselines under limited search budgets (N=20 and N=30).
  • The training dynamics show a clear downward trend in loss for both movement and placement modules, indicating effective optimization.

Limitations and Future Work

  • Training Stability: The placement module's optimization process exhibits more fluctuations than the movement module, suggesting potential areas for improvement in training stability.
  • Decision Strategy: The current final decision strategy relies on the overall objective value with random tie-breaking. The authors suggest developing a more principled decision strategy for further improvement.
  • Rigorous Training Criteria: Future work will investigate more rigorous criteria for assessing training completion and convergence, as loss values alone may not fully reflect model maturity.
  • Absolute Strength: While the framework is competitive and resource-efficient, it does not surpass established handcrafted engines like "Invader" in absolute playing strength, which are built on years of domain-specific engineering. The primary goal was to demonstrate effective learning from weak supervision in resource-constrained settings.

Summary updated

BibTeX
@misc{qian2026resource,
  author = {Tianhao Qian and Zhuoxuan Li and Jinde Cao and Xinli Shi and Hanjie Liu and Leszek Rutkowski},
  title = {{Resource-constrained Amazons chess decision framework integrating large language models and graph attention}},
  year = {2026},
  eprint = {2603.10512},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2603.10512},
}
Figure