Google DeepMind Paper Introduces SkillSmith for Parametric Skills

Posts describe new Google DeepMind research called SkillSmith. The method treats model weights as a native input modality the LLM processes directly. It ingests prefix weights alongside text that explains how learned capabilities relate to a goal, then generates fresh prefix weights that realize the skill in one forward pass. The posts note the model can now draw on both written instructions and prior weight-based knowledge instead of starting from scratch or averaging weights without guidance.