AI Paper Atlas

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

ICLR 2026 · Large language models › Reasoning and chain of thought · approach: Routing, pruning and distillation for reasoning

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 changedWhereMeasuredEffect
port refusal adapter from 3B to 8B via Command-V instead of training on 8B directlyLlama3.1-8B-Instruct · WildJailbreak adversarial harmfulattack success rate: 28.55% (direct ReFT) vs 20.40% (Command-V)Improved
port refusal adapter from 8B to 3B via Command-VLlama3.2-3B-Instruct · WildJailbreak adversarial harmfulattack success rate: 31.30% -> 16.4%Improved
transfer refusal adapter with Command-V instead of ReFT trainingLlama 3.1 8B · WildJailbreak adversarial benignoverrefusal rate: 2.26 (ReFT) -> 4.52 (Command-V); off-shelf 1.36Hurt
add refusal adapters (ReFT or Command-V)Llama 3.1 8B / Llama 3.2 3B · ARC-Challengeaccuracy: 85.75 -> 86.43 (8B); 80.12 -> 78.24 (3B)No clear effect
port jailbreak adapters across models with Command-Vinstruct models (Llama3, Qwen2.5, Gemma2) · AdvBench / HarmBenchattack success rate: single-digit % -> 20-80% for many; Gemma2-2B-it resistsMixed
port CoT adapter via Command-V from same-family modelQwen2.5 / Llama3 models · BBH Multistep Arithmeticaccuracy: +50% for Qwen 2.5 3BImproved
port CoT adapter via Command-VQwen2.5 / Llama3 models · BBH Tracking Shuffled Objects (five)accuracyMixed
port CoT adapter via Command-Vrecipient models · JSON format compliancecompliant responses: up to 28.8% of responses non-compliantHurt
increase profiling prompts N from 50 to 1000 (ablation)Command-V converters (8B to 3B) · converter test MSEforward test MSE (1e-5): 2.1 -> 1.5Improved
use more diverse mixed profiling prompts (N=2000) (ablation)Command-V converters (8B to 3B) · converter test MSEforward test MSE (1e-5): 1.5 -> 0.5Improved
use MSE-minimizing layer matching instead of linear depth mapping (ablation)Command-V layer correspondence · downstream taskstransfer qualityHurt
apply Command-V to adapters with little donor effect (commonsense reasoning)recipient models · commonsense reasoningperformanceNo clear effect

Experimental setup

Builds on: representation finetuning

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy