UniMaia: Language-Guided Chess Policy for Human-Like Play
A team of researchers has introduced UniMaia, a novel framework that leverages natural language prompts to manipulate a static Lc0-based chess policy network. This is achieved through a text encoder that is efficient in terms of parameters and employs ControlNet-style conditioning. The framework allows for semantic control of various aspects of the game, such as choosing openings and adjusting player strength, while preserving the essential domain-specific inductive biases. Additionally, an auxiliary version known as UniMaia-Aux features temporal conditioning. Overall, this method strikes a balance between flexibility and effectiveness in structured decision-making.
Key facts
- UniMaia adapts a frozen Lc0-based chess policy network using a text encoder and ControlNet-style conditioning.
- It enables semantic control over opening selection and player strength.
- UniMaia-Aux incorporates auxiliary temporal conditioning.
- The framework preserves pretrained policy representations.
- It addresses the trade-off between flexibility and domain-specific performance.
- Large language models serve as a flexible interface for controlling complex systems.
- Specialized policy networks lack semantic controllability.
- Prompt-conditioned language models exhibit weaker domain grounding.
Entities
—