Mixture of Masters: Merging Specialized Chess Language Models

Recent advancements in Language Models (LMs) have raised the possibility of applying these models to strategic domains beyond text, such as chess. A significant challenge in this area is determining whether LMs can replicate the distinctive playing styles of an individual, rather than merely generating plausible moves. Addressing this challenge involves training models that are not only accurate in move generation but also capable of capturing stylistic nuances unique to each player. To tackle this problem, we propose a *Mixture of Experts* (MoE) framework for chess, where each expert model is fine-tuned separately on the historical games of specific grandmasters, including Hikaru Nakamura, Magnus Carlsen, and Garry Kasparov. This approach allows each model to specialize in a single grandmaster's style, and we then explore techniques for merging these experts into a combined model that retains the stylistic diversity of its individual components. To evaluate the efficacy of this MoE framework, we developed a custom evaluation system incorporating metrics for move legality, style fidelity, and performance benchmarking against the Stockfish engine. Our results show that the expert models achieve up to 16.7% more accuracy in replicating grandmaster moves compared to the baseline, with a high rate of legal move generation. The merged model, while experiencing a slight decrease in accuracy, exhibits emergent properties, demonstrating the potential for further improvements in model merging techniques. This research advances our understanding of specialization and model merging in LMs, highlighting their potential for complex strategic games and human-AI alignment.