UniMaia: Steering Chess Policies with Language for Human-like Play

UniMaia is a new AI framework that uses natural language prompts to control chess-playing policy networks, enabling semantic control over gameplay elements like opening selection and player strength without requiring large-scale multimodal training. The system combines a frozen Lc0 chess engine with a parameter-efficient text encoder and demonstrates competitive performance on prompt-conditioned benchmarks while maintaining domain-specific expertise.

UniMaia addresses a fundamental challenge in AI systems: balancing domain-specific performance with flexible semantic control. Traditional specialized policy networks excel at chess but lack interpretability and controllability, while large language models offer flexibility at the cost of weak domain grounding. This research bridges that gap through parameter-efficient adaptation, allowing natural language to steer a pretrained chess engine without degrading its core capabilities.

The framework's innovation lies in its architectural elegance. By freezing the underlying Lc0 policy network and adding only a lightweight text encoder with ControlNet-style conditioning, UniMaia avoids expensive end-to-end retraining. The introduction of UniMaia-Aux with auxiliary objectives for temporal conditioning and behavioral prediction further refines controllability while maintaining competitive performance. The team's construction of a large-scale metadata-augmented Lichess dataset and semi-automated prompt-generation pipeline demonstrates methodological rigor.

Beyond chess, this work signals broader implications for AI systems. The approach demonstrates that domain expertise and linguistic flexibility aren't mutually exclusive when properly architected. This pattern could extend to other structured domains—trading systems, game engines, scientific modeling—where combining specialized knowledge with natural language control creates more interpretable and steerable AI systems.

The trade-offs highlighted between controllability and predictive performance are realistic and important. Developers seeking to apply similar approaches should expect similar compromises. The competitive performance across multiple benchmark categories suggests the framework generalizes reasonably well, making it a template for future work combining language models with domain-specific networks.

• →UniMaia enables natural language control of chess policy networks through parameter-efficient adaptation rather than end-to-end training.
• →The framework achieves state-of-the-art results on prompt-conditioned benchmarks while maintaining domain-specific performance without multimodal training.
• →UniMaia-Aux demonstrates that auxiliary conditioning objectives improve controllability with acceptable trade-offs in certain metrics.
• →Parameter-efficient text encoders and ControlNet-style mechanisms offer a scalable pattern for adding semantic control to frozen domain-specific models.
• →The approach suggests broader applicability beyond chess to other structured decision-making domains requiring both expertise and interpretability.