elvis

Rohan Paul

A new paper from Google DeepMind introduces SkillSmith, a system that treats model weights as something an LLM can read and act on, not just a static artifact to be fine-tuned or averaged.
The core idea: for a new task, SkillSmith feeds a frozen Gemma 3 4B model both existing prefix key-value caches and a text description explaining how known capabilities relate to the target skill. A single forward pass then outputs a fresh prefix cache that encodes the new capability. The team calls this “instruction-steered parametric synthesis.”
The practical upshot is that skill composition becomes an inference-time operation rather than a new training run.
The ablation results make the case clearly. On the Composite-SNI benchmark, using only K-V caches gives an Elo of 1,455. Text-only gets 1,622. Combining both reaches 1,714. Neither richer prompting nor a learned weight merger alone explains that gap.
SkillSmith also gives downstream fine-tuning a better starting point. On data-sparse MMLU-ProX tasks, the generated prefixes produce stronger results than starting from scratch. That said, on simpler SNI tasks with around 1,000 examples each, standard fine-tuned methods catch up.
So SkillSmith doesn’t eliminate training. What it changes is where adaptation begins: previous text and previous weights can now be composed into the initialization for a new capability.
Paper: <a href="https://arxiv.org/abs/2607.27497" target="_blank" rel="noopener nofollow">SkillSmith: Learning to Compose Parametric Skills and Textual Knowledge</a>

Collections

Google DeepMind's SkillSmith is a new system that composes LLM capabilities at inference time rather than through retraining. It feeds a frozen Gemma 3 4B model both existing prefix key-value caches and textual descriptions of known skills, producing a new prefix cache encoding a target capability in a single forward pass — a process called 'instruction-steered parametric synthesis.' On the Composite-SNI benchmark, combining KV caches and text descriptions achieves an Elo of 1,714, outperforming either modality alone. SkillSmith also provides better initialization for downstream fine-tuning on data-sparse tasks like MMLU-ProX, though standard fine-tuning catches up on simpler tasks with sufficient data. The key shift is treating model weights as readable inputs, making skill composition an inference-time operation rather than a training-time one.

Google DeepMind's SkillSmith composes model skills at inference time by treating weights as input