Research map / Large language models / Reasoning and chain of thought
Command-V: Training-Free Representation Finetuning Transfer
Barry Wang, Avi Schwarzschild, Alexander Robey, Ali Payani, Charles Fleming, Mingjie Sun, Daphne Ippolito
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Command-V (⌘V), a backpropagation-free method that transfers a residual representation adapter from a donor LLM into an architecturally different recipient model.
Problem
Retrofitting LLMs with new behaviors usually requires full finetuning or distillation, which is costly and must be repeated for every architecture.
Approach
Copies an existing residual representation adapter from a donor model and pastes its effect into an architecturally different recipient model without backpropagation, enabling cross-architecture behavior transfer.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| port refusal adapter from 3B to 8B via Command-V instead of training on 8B directly | Llama3.1-8B-Instruct · WildJailbreak adversarial harmful | attack success rate: 28.55% (direct ReFT) vs 20.40% (Command-V) | Improved |
| port refusal adapter from 8B to 3B via Command-V | Llama3.2-3B-Instruct · WildJailbreak adversarial harmful | attack success rate: 31.30% -> 16.4% | Improved |
| transfer refusal adapter with Command-V instead of ReFT training | Llama 3.1 8B · WildJailbreak adversarial benign | overrefusal rate: 2.26 (ReFT) -> 4.52 (Command-V); off-shelf 1.36 | Hurt |
| add refusal adapters (ReFT or Command-V) | Llama 3.1 8B / Llama 3.2 3B · ARC-Challenge | accuracy: 85.75 -> 86.43 (8B); 80.12 -> 78.24 (3B) | No clear effect |
| port jailbreak adapters across models with Command-V | instruct models (Llama3, Qwen2.5, Gemma2) · AdvBench / HarmBench | attack success rate: single-digit % -> 20-80% for many; Gemma2-2B-it resists | Mixed |
| port CoT adapter via Command-V from same-family model | Qwen2.5 / Llama3 models · BBH Multistep Arithmetic | accuracy: +50% for Qwen 2.5 3B | Improved |
| port CoT adapter via Command-V | Qwen2.5 / Llama3 models · BBH Tracking Shuffled Objects (five) | accuracy | Mixed |
| port CoT adapter via Command-V | recipient models · JSON format compliance | compliant responses: up to 28.8% of responses non-compliant | Hurt |
| increase profiling prompts N from 50 to 1000 (ablation) | Command-V converters (8B to 3B) · converter test MSE | forward test MSE (1e-5): 2.1 -> 1.5 | Improved |
| use more diverse mixed profiling prompts (N=2000) (ablation) | Command-V converters (8B to 3B) · converter test MSE | forward test MSE (1e-5): 1.5 -> 0.5 | Improved |
| use MSE-minimizing layer matching instead of linear depth mapping (ablation) | Command-V layer correspondence · downstream tasks | transfer quality | Hurt |
| apply Command-V to adapters with little donor effect (commonsense reasoning) | recipient models · commonsense reasoning | performance | No clear effect |
Experimental setup
Builds on: representation finetuning
Closest papers by meaning
- Model-Aware Tokenizer Transfer (NeurIPS 2026 · similarity 0.87)
- Memory Grafting: Scaling Language Model Pre-training via Offline Conditional Memory (NeurIPS 2026 · similarity 0.87)
- Cross-model Control: Improving Multiple Large Language Models in One-time Training (NeurIPS 2024 · similarity 0.86)
- Neutral residues: revisiting adapters for model extension (ICML 2025 · similarity 0.86)
- Drag-and-Drop LLMs: Zero-Shot Prompt-to-Weights (NeurIPS 2025 · similarity 0.86)
- ASRU: Activation Steering Meets Reinforcement Unlearning for Multimodal Large Language Models (ICML 2026 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
