Enregistré dans:
Détails bibliographiques
Auteurs principaux: Wang, Barry, Schwarzschild, Avi, Robey, Alexander, Payani, Ali, Fleming, Charles, Sun, Mingjie, Ippolito, Daphne
Format: Preprint
Publié: 2025
Sujets:
Accès en ligne:https://arxiv.org/abs/2506.19140
Tags: Ajouter un tag
Pas de tags, Soyez le premier à ajouter un tag!
_version_ 1866915357392896000
author Wang, Barry
Schwarzschild, Avi
Robey, Alexander
Payani, Ali
Fleming, Charles
Sun, Mingjie
Ippolito, Daphne
author_facet Wang, Barry
Schwarzschild, Avi
Robey, Alexander
Payani, Ali
Fleming, Charles
Sun, Mingjie
Ippolito, Daphne
contents Retrofitting large language models (LLMs) with new behaviors typically requires full finetuning or distillation-costly steps that must be repeated for every architecture. In this work, we introduce Command-V, a backpropagation-free behavior transfer method that copies an existing residual activation adapter from a donor model and pastes its effect into a recipient model. Command-V profiles layer activations on a small prompt set, derives linear converters between corresponding layers, and applies the donor intervention in the recipient's activation space. This process does not require access to the original training data and needs minimal compute. In three case studies-safety-refusal enhancement, jailbreak facilitation, and automatic chain-of-thought reasoning--Command-V matches or exceeds the performance of direct finetuning while using orders of magnitude less compute. Our code and data are accessible at https://github.com/GithuBarry/Command-V/.
format Preprint
id arxiv_https___arxiv_org_abs_2506_19140
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Command-V: Pasting LLM Behaviors via Activation Profiles
Wang, Barry
Schwarzschild, Avi
Robey, Alexander
Payani, Ali
Fleming, Charles
Sun, Mingjie
Ippolito, Daphne
Machine Learning
Retrofitting large language models (LLMs) with new behaviors typically requires full finetuning or distillation-costly steps that must be repeated for every architecture. In this work, we introduce Command-V, a backpropagation-free behavior transfer method that copies an existing residual activation adapter from a donor model and pastes its effect into a recipient model. Command-V profiles layer activations on a small prompt set, derives linear converters between corresponding layers, and applies the donor intervention in the recipient's activation space. This process does not require access to the original training data and needs minimal compute. In three case studies-safety-refusal enhancement, jailbreak facilitation, and automatic chain-of-thought reasoning--Command-V matches or exceeds the performance of direct finetuning while using orders of magnitude less compute. Our code and data are accessible at https://github.com/GithuBarry/Command-V/.
title Command-V: Pasting LLM Behaviors via Activation Profiles
topic Machine Learning
url https://arxiv.org/abs/2506.19140