OpenNovelty.org

Command-V: Training-Free Representation Finetuning Transfer

ICLR 2026 Conference SubmissionAnonymous Authors
OpenReview Score: 7.0Download Report PDF
Model EditingSteeringActivationInterpretability
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 ⌘V (Command-V), a backpropagation-free behavior transfer method that copies an existing residual representation adapter from a donor model and pastes its effect into an architecturally different recipient model. ⌘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—⌘V matches the performance of direct finetuning while using orders of magnitude less resources.

Disclaimer
This report is AI-GENERATED using Large Language Models and WisPaper (A scholar search engine). It analyzes academic papers' tasks and contributions against retrieved prior work. While this system identifies POTENTIAL overlaps and novel directions, ITS COVERAGE IS NOT EXHAUSTIVE AND JUDGMENTS ARE APPROXIMATE. These results are intended to assist human reviewers and SHOULD NOT be relied upon as a definitive verdict on novelty.
NOTE that some papers exist in multiple, slightly different versions (e.g., with different titles or URLs). The system may retrieve several versions of the same underlying work. The current automated pipeline does not reliably align or distinguish these cases, so human reviewers will need to disambiguate them manually.
If you have any questions, please contact: mingzhang23@m.fudan.edu.cn

Overview

Overall Novelty Assessment

The paper introduces Command-V, a method for transferring learned behaviors between architecturally different language models by mapping activation spaces without backpropagation. It resides in the 'Activation-Space Representation Transfer' leaf, which contains only one sibling paper (Activation Manifold Projection). This leaf sits within the broader 'Cross-Architecture Adapter and Module Transfer' branch, which includes three leaves total. The sparse population suggests this is an emerging rather than saturated research direction, with relatively few prior works directly addressing activation-space behavior transfer across heterogeneous architectures.

The taxonomy reveals that neighboring approaches tackle similar cross-architecture challenges through different mechanisms. The sibling leaf 'Low-Rank Adapter Transfer' focuses on projecting LoRA modules rather than activation representations, while 'Linear-Cost Architecture Transfer' addresses migration to state-space models specifically. Adjacent branches explore prompt-based methods and memory augmentation, which avoid weight-space interventions entirely. Command-V occupies a middle ground: it manipulates internal representations like prompt-based methods but operates through learned linear converters rather than external memory or discrete prompts, distinguishing it from both adapter projection and pure inference-time techniques.

Among 25 candidates examined across three contributions, none were flagged as clearly refuting the work. The 'activation profiling method' examined 7 candidates with no refutations, the 'Command-V adapter transfer framework' examined 8 with none, and the 'training-free behavior transfer method' examined 10 with none. This suggests that within the limited search scope—focused on top-K semantic matches and citation expansion—no prior work was found that directly anticipates the combination of activation profiling, linear conversion, and cross-architecture adapter transfer. The statistics indicate a relatively clean novelty signal, though the modest search scale (25 papers) means exhaustive coverage cannot be claimed.

Based on the limited literature search, Command-V appears to occupy a sparsely populated niche within activation-space transfer. The absence of refuting candidates across all contributions, combined with the leaf's small sibling count, suggests the specific approach is not directly anticipated by examined prior work. However, the search scope of 25 papers leaves open the possibility that related techniques exist in adjacent subfields or recent preprints not captured by semantic search.

Taxonomy

Core-task Taxonomy Papers
46
3
Claimed Contributions
25
Contribution Candidate Papers Compared
0
Refutable Paper

Research Landscape Overview

Core task: training-free behavior transfer across different language model architectures. The field addresses how to move learned capabilities—ranging from task-specific adaptations to entire behavioral policies—from one model architecture to another without retraining from scratch. The taxonomy reveals eight major branches that span diverse transfer mechanisms. Cross-Architecture Adapter and Module Transfer focuses on moving lightweight components such as LoRA[1] modules or activation-space representations between models with different internal structures. Cross-Lingual and Multilingual Transfer examines how linguistic knowledge propagates across language boundaries, often leveraging shared representations or tokenizer mappings like Zero-shot Tokenizer Transfer[9]. Prompt-Based and Memory-Augmented Transfer explores training-free methods that rely on external memory or carefully designed prompts to guide behavior. Domain and Task Adaptation Without Training tackles specialized settings—medical QA, robotics, or financial forecasting—where direct fine-tuning is impractical. Behavioral and Reinforcement Learning Transfer investigates policy or decision-making transfer, including works like LLMs to RL[33] that bridge language models and sequential decision tasks. Specialized Transfer and Distillation Methods cover techniques such as pruning, knowledge diversion, and meta-learning that compress or reshape models. Foundation Model Composition and Unified Frameworks study how to combine multiple pretrained models into coherent systems, exemplified by Socratic Models[8]. Finally, Emerging and Experimental Transfer Paradigms capture novel directions that do not yet fit established categories. A particularly active line of work centers on activation-space and module-level transfer, where researchers seek to align internal representations or adapter weights across architectures with minimal overhead. Command-V[0] exemplifies this direction by transferring behavior through activation-space mappings, closely related to Activation Manifold Projection[23], which also manipulates hidden states to achieve cross-architecture compatibility. These approaches contrast with heavier distillation pipelines or prompt-engineering strategies, offering a middle ground between full retraining and purely inference-time interventions. Meanwhile, cross-lingual studies like Cross-lingual Transfer Factors[7] and behavioral policy transfer methods such as Behavioral Foundation Models[2] highlight alternative pathways: the former emphasizes linguistic structure, while the latter focuses on decision-making patterns. Open questions remain around scalability—whether activation-space methods generalize to very large or highly dissimilar architectures—and the trade-offs between transfer fidelity and computational cost. Command-V[0] sits squarely within the activation-space cluster, sharing conceptual ground with Activation Manifold Projection[23] but differing in how it handles architectural mismatches and the granularity of behavior it aims to preserve.

Claimed Contributions

Activation profiling method

The authors introduce activation profiling, a technique that records and analyzes layer activations from a small set of prompts to identify corresponding activation patterns between different transformer-based language models, enabling cross-model behavior transfer without requiring architectural similarity.

7 retrieved papers
Command-V adapter transfer framework

The authors develop Command-V, a training-free framework that uses activation profiles to derive linear converters between model layers, allowing representation adapter weights from a donor model to be transferred and applied to an architecturally different recipient model without backpropagation or additional training data.

8 retrieved papers
Training-free behavior transfer method

The authors present Command-V as a complete method that transfers learned behaviors across different model architectures by profiling activations, deriving converters, and applying donor interventions in the recipient's activation space, requiring minimal compute and no access to original training data.

10 retrieved papers

Core Task Comparisons

Comparisons with papers in the same taxonomy category

Contribution Analysis

Detailed comparisons for each claimed contribution

Contribution

Activation profiling method

The authors introduce activation profiling, a technique that records and analyzes layer activations from a small set of prompts to identify corresponding activation patterns between different transformer-based language models, enabling cross-model behavior transfer without requiring architectural similarity.

Contribution

Command-V adapter transfer framework

The authors develop Command-V, a training-free framework that uses activation profiles to derive linear converters between model layers, allowing representation adapter weights from a donor model to be transferred and applied to an architecturally different recipient model without backpropagation or additional training data.

Contribution

Training-free behavior transfer method

The authors present Command-V as a complete method that transfers learned behaviors across different model architectures by profiling activations, deriving converters, and applying donor interventions in the recipient's activation space, requiring minimal compute and no access to original training data.

Table of Contents

Command-V: Training-Free Representation Finetuning Transfer | Novelty Validation