analiza aiqbench
This paper proposes CUSD, a data-free method that splits fine-tuned models' task vectors into a shared directional component and a task-specific tail before merging. Its reported gains (improvements in 42 of 44 model-operator combinations, up to 22.5%) are substantial but come from benchmarks across specific base models and arithmetic operators, so practitioners should check how well they transfer to their own setups. For ML engineers, the main takeaway is that treating all task-vector directions with one element-wise rule may leave performance on the table, and that the common-versus-unique split is a cheap pre-processing step that requires no training data.
Separating Common and Unique Directions for Data-Free Model Merging
arXiv:2611.04108v1 Announce Type: new Abstract: Model merging combines multiple fine-tuned models without additional training, and arithmetic methods have emerged as a simple yet effective way to merge models without data. However, existing arithmetic methods treat task vectors in parameter space and apply a single element-wise rule to directions that can differ in mergeability. In a shared directional basis, we find a consistent two-regime structure: a low-dispersion common bulk shared across tasks and a high-dispersion task-dominated tail that carries disproportionate energy. These findings motivate Common--Unique Subspace Decomposition (CUSD), a data-free pre-merging method that separates task vectors into common and unique directional components, merges the common components with downstream arithmetic operator, and retains unique residuals separately. Experiments across $11$ base models, four model families, and four arithmetic-based merging operators show that CUSD improves 42 of 44 model--operator combinations and achieves up to 22.5% improvement over the corresponding baselines. Our code is publicly available at https://github.com/stwang666/CUSD
Abstract:Model merging combines multiple fine-tuned models without additional training, and arithmetic methods have emerged as a simple yet effective way to merge models without data. However, existing arithmetic methods treat task vectors in parameter space and apply a single element-wise rule to directions that can differ in mergeability. In a shared directional basis, we find a consistent two-regime structure: a low-dispersion common bulk shared across tasks and a high-dispersion task-dominated tail that carries disproportionate energy. These findings motivate Common--Unique Subspace Decomposition (CUSD), a data-free pre-merging method that separates task vectors into common and unique directional components, merges the common components with downstream arithmetic operator, and retains unique residuals separately. Experiments across $11$ base models, four model families, and four arithmetic-based merging operators show that CUSD improves 42 of 44 model--operator combinations and achieves up to 22.5% improvement over the corresponding baselines. Our code is publicly available at this https URL
Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
arXiv:2611.04108 [cs.LG]
(or arXiv:2611.04108v1 [cs.LG] for this version)
https://doi.org/10.48550/arXiv.2611.04108 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Text integral sub licența deschisă a publicației · 247 cuvinte
Subiecte conexe
Can the Assistant Axis Predict Reward-Hacking?
LessWrong · 26 oct. 2026
Directional Linear Separability of Neural Representations: Geometry and Transformations
arXiv · 22 sept. 2026
Skin-Deep: A Geometric Diagnostic for Alignment Fragility in Large Language Model Representations
arXiv · 2 oct. 2026
Empathy Is Steerable but Multi-Axial: Mechanism Geometry and Persona Effects in LLMs
arXiv · 15 sept. 2026