Command-V: Pasting LLM Behaviors via Activation Profiles

2025-06-23

AI Summary

How to efficiently transfer novel capabilities from a donor large language model (LLM) to a recipient LLM without fine-tuning, knowledge distillation, or access to the original training data—especially across architecturally distinct models? Method: We propose a behavior transplantation framework that analyzes layer-wise activation profiles under small-scale prompting to learn a linear residual activation transformer between source and target models; the resulting adapter module is then directly inserted (“copy-paste”) into the recipient’s forward pass. Contribution/Results: This is the first method enabling zero-backpropagation, zero-data-dependency cross-architecture behavioral transfer. It achieves performance on par with or exceeding full fine-tuning on three challenging tasks—safety refusal, jailbreak defense, and automatic chain-of-thought generation—while reducing computational overhead by 1–2 orders of magnitude, thereby significantly enhancing the efficiency of LLM capability reuse.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: (Large) Language Models
• Computer Vision: Large Vision Models

Application Category

• Search and Retrieval-Augmented AI: Search Tool Learning with LLM: Teaching LLMs to invoke search and make use of retrieved information
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions

Abstract

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/.

Problem

• Transferring LLM behaviors without full finetuning
• Enhancing model performance with minimal compute
• Applying donor interventions in recipient activation space

Innovation

• Backpropagation-free behavior transfer method
• Linear converters between corresponding layers
• Minimal compute without original training data