DeepMind's SkillSmith Treats Model Weights as a Modality for Inference-Time Skill Composition

omarsar0 · x · 2026-08-03

Google DeepMind introduced SkillSmith, a novel approach that treats model weights as an additional modality natively read by LLMs.

The augmented model ingests text describing how a capability relates to a target and directly outputs new prefix weights manifesting that skill. This transforms skill composition from a training run into an inference-time operation, which the team calls "instruction-steered parametric synthesis."

Experiments show that combining text and weight adaptation yields gains exceeding what text-only or weight-only methods can achieve independently.

Original post →

More from Research

Research channel →