Fine-tuning a base model for each task gives separate checkpoints. Merging them into one model without original data is efficient but the changes interfere. Existing methods treat each change individually, missing how they interact inside the model.
Joint Null-Space Projection first identifies the key internal directions each model adjusts. It then removes from each model any changes that overlap with important directions of other models, preserving unique knowledge without conflict. This requires no data.
On eight image-classification tasks, it achieved 89.45% average accuracy, outperforming previous data-free merging by 11.88 points. This shows that resolving interference at the directional level improves merging. It did not work on language tasks, highlighting limitations.