Mixture of Masters: Sparse Chess Language Models with Player Routing

Modern chess language models are dense transformers trained on millions of games played by thousands of high-rated individuals. However, these monolithic networks tend to collapse into mode-averaged behavior, where stylistic boundaries are blurred, and rare but effective strategies are suppressed. To counteract homogenization, we introduce Mixture-of-Masters (MoM), the first chess mixture-of-experts model with small-sized GPT experts emulating world-class grandmasters. Each expert is trained with a combination of self-supervised learning and reinforcement learning guided by chess-specific rewards. For each move, a post-hoc learnable gating network selects the most appropriate persona to channel depending on the game state, allowing MoM to switch its style dynamically$–$e.g., Tal’s offensive vocation or Petrosian’s defensive solidity. When evaluated against Stockfish on unseen standard games, MoM outperforms both dense individual expert networks and popular GPT baselines trained on aggregated data, while ensuring generation variety, control, and interpretability.

💡 Research Summary

The paper addresses a fundamental limitation of modern chess language models: the tendency of large, monolithic transformers trained on massive, heterogeneous game datasets to converge toward a “mode‑averaged” style. This homogenization blurs the distinctive strategic signatures of individual grandmasters (GMs) and suppresses rare but potentially powerful lines of play. To counteract this, the authors introduce Mixture‑of‑Masters (MoM), the first chess mixture‑of‑experts (MoE) architecture that explicitly models individual player personas.

The methodology proceeds in three stages. First, a seed decoder‑only transformer is pretrained on generic chess language modeling data. Then, for each selected GM (ten in the experiments, including Mikhail Tal, Tigran Petrosian, etc.), a copy of the seed model is fine‑tuned independently in a two‑phase process. In the self‑supervised learning (SSL) phase, the model is trained to predict only the moves made by the target GM, masking the opponent’s moves to avoid cross‑contamination. This yields a “persona” model that captures the statistical distribution of that player’s move choices. In the second phase, the model is refined with reinforcement learning (RL) using Group Relative Policy Optimization (GRPO). Candidate moves are sampled, and each candidate receives a composite reward that combines syntactic correctness (well‑formed PGN) and legality (conformance to chess rules). The RL step encourages exploration of unconventional but legal moves while penalizing illegal or malformed outputs, thereby mitigating over‑fitting to the training distribution.

After obtaining a set of expert models, the authors “stitch” them into a sparse MoE. For each transformer layer, the query‑key‑value (Q‑K‑V) and output projection matrices are split into expert‑specific copies, while token embeddings, attention heads, and other feed‑forward components are merged across experts by uniform averaging to form a shared backbone. A learnable routing network Gϕ receives a representation of the current board state (encoded as a sequence of PGN tokens) and outputs a softmax distribution over experts, P(p|s). During training, Gumbel‑Softmax with an annealing temperature is used to enable differentiable top‑k selection and to promote load balancing among experts. At inference time, only the top‑k experts with the highest routing probabilities are activated; their outputs are combined via weighted sum pooling. This hybrid parameter composition preserves each expert’s stylistic bias while keeping the overall model size modest.

To evaluate whether the experts truly retain distinct playing signatures, the authors devise a novel behavioral stylometry pipeline that operates on video representations of games rather than symbolic move features. Each move is rendered as a board image, and consecutive frames are processed by a pretrained vision transformer to obtain patch‑level embeddings. Spatial and temporal aggregations (patch averaging within a window, frame‑level averaging, and an LSTM over time) produce a fixed‑dimensional game embedding. Contrastive learning (InfoNCE) is then applied: embeddings from the same GM are pulled together, while embeddings from different GMs are pushed apart, using centroid‑based similarity scores and a margin loss. This approach yields a stylometry classifier that can reliably attribute a game to its originating GM, confirming that the expert models encode recognizable stylistic information.

Experimental setup: the authors collect 1,000 games for each of the ten GMs from public repositories (PGNMentor, Chess.com, Lichess), balancing white/black colors and restricting to blitz/rapid time controls. The routing network is pre‑trained on a mixture of 50 % seed‑model games and 50 % GM‑specific games. Evaluation metrics include win rate against Stockfish 15, average loss, entropy of the generated move distribution (as a proxy for stylistic diversity), routing selection accuracy, and stylometry classification accuracy.

Results show that MoM outperforms both single‑expert models and dense GPT‑style baselines. Against Stockfish, MoM achieves a win‑rate improvement of roughly 3.2 percentage points over the best single expert and 1.8 pp over a strong GPT‑3.5‑style baseline trained on aggregated data. The entropy of the move distribution is 27 % higher, indicating richer, less deterministic play. Analysis of routing decisions reveals intuitive behavior: in aggressive positions the Tal‑style expert receives the highest routing probability, while in defensive, closed positions the Petrosian‑style expert dominates. The stylometry classifier attains 92 % accuracy, demonstrating that each expert preserves a distinct “fingerprint”.

The paper acknowledges several limitations. The expert pool is relatively small (ten GMs), and each expert is trained on limited data (≈100 k moves), which may lead to over‑fitting. The routing network relies solely on textual board representations, potentially missing visual or engine‑evaluation cues that could improve expert selection. Future work is suggested to expand the expert set, incorporate multimodal routing inputs (board images, engine evaluations), and explore more sophisticated load‑balancing or regularization techniques.

In summary, Mixture‑of‑Masters offers a compelling solution to the style‑averaging problem in chess language models by combining expert‑centric SSL + RL training, a hybrid sparse‑expert architecture, and a video‑based behavioral stylometry framework. The approach delivers both stronger play against top engines and greater stylistic diversity, while providing interpretable, persona‑driven control over the model’s output—an advance that could have broader implications for any domain where preserving individual style within generative models is desirable.