SkillSmith: Learning to Compose Parametric Skills and Textual Knowledge

Lucio M. Dery*, Benedict Aaron Tjandra*, Siavash Samiei, Adhiguna Kuncoro, Zohar Yahav, Jiajun Shen, and Arthur Szlam

* Equal contribution.

COLM 2026 LLA Workshop

SkillSmith combines parametric skills with textual knowledge to produce new capabilities. Given learned prefix weights, text, and a target instruction, the model generates new prefix weights for the desired skill. This brings skill composition and text-based adaptation into a shared framework, outperforming text-only and weight-only baselines in our experiments.