Command-V: Pasting LLM Behaviors via Activation Profiles
This addresses the need for efficient behavior adaptation in LLMs, offering a low-cost alternative to traditional methods, though it is incremental as it builds on existing activation-based techniques.
The authors tackled the problem of retrofitting large language models with new behaviors without costly full finetuning or distillation by introducing Command-V, a backpropagation-free method that transfers behaviors via activation profiles, achieving performance matching or exceeding direct finetuning with orders of magnitude less compute in case studies like safety-refusal enhancement.
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/.