DeepMind's SkillSmith: Transforming Model Weights into Composable Skill Modules

eyishazyer · x · 2026-08-03

A new paper from Google DeepMind suggests that model weights do not have to remain static artifacts that can only be fine-tuned or averaged.

The proposed SkillSmith framework treats prefix key-value caches as another input modality. By interleaving existing prefix weights with task descriptions and examples, the system can initialize new capabilities from what the agent previously learned—marking a shift from learning isolated skills to composing them.

Related event: DeepMind Proposes SkillSmith: Treating Model Weights as Composable Modalities(4 posts)→

Original post →

More from Research

Research channel →