A new paper from Google DeepMind introduces “SkillSmith,” a method that enhances the initialization of model capabilities by combining both written outputs from an agent and previously learned model weights. This approach allows for the composition of skills at inference time, eliminating the need for a separate training run, and it generates prefixes that serve as stronger starting points for fine-tuning on difficult tasks. The findings, demonstrated through various benchmark tests, indicate that SkillSmith significantly improves performance, particularly on challenging, data-sparse tasks, by interleaving existing knowledge with task-specific information.

Gemma 3: Gemma 3 is a family of language models developed by Google DeepMind. SkillSmith applies its composition method to a frozen Gemma 3 4B model, generating fresh prefix caches from combined inputs without further training on the target task.
SkillSmith: SkillSmith is a method for learning to compose parametric skills and textual knowledge into new model capabilities. It treats prefix key-value caches as an additional input modality that can be interleaved with task descriptions and rationales. The technique is presented in a new Google DeepMind paper and allows composition at inference time rather than through dedicated training.
Google DeepMind: Google DeepMind is an AI research organization that develops advanced machine learning models and techniques. It published the paper introducing SkillSmith, which enables new ways to compose model capabilities from both learned weights and textual knowledge. The work demonstrates this approach using one of its models in a frozen state for inference-time adaptation.
SkillSmith: Learning to Compose Parametric Skills and Textual Knowledge: This is the title of the arXiv paper from Google DeepMind that introduces the SkillSmith method. It details how previous model weights and text can be combined to initialize new capabilities for a target task.

Composition Method: SkillSmith enables new capabilities to be initialized from both written agent outputs and previously learned model weights.
Fine-tuning Support: Prefixes generated by SkillSmith provide stronger starting points for subsequent fine-tuning on challenging tasks.
Inference Adaptation: The approach performs initial composition of skills at inference time rather than requiring a new target-specific training run.