
LLM-Graph Hybrid Beats GPT-4o-mini in Amazons
Proposes a lightweight hybrid framework for resource-constrained Amazons chess, combining graph attention autoencoders with MCTS, genetic algorithms, and GPT-4o-mini for synthetic data. Achieves 15%-56% higher decision accuracy over baselines and outperforms GPT-4o-mini with 66.5% win rate at N=50 nodes. Demonstrates weak-to-strong generalization from noisy LLM supervision.





