Merge++ improves multi-task model merging without extra data
Merge++: Universal Merge Refinement Through Data-Free Checkpoint Inversion
Machine LearningArtificial Intelligence
Summary
Merging separate AI models trained for different tasks into one model usually requires data that might not be available. Existing methods only mix the models’ internal settings without checking how each model behaves. The authors propose Merge++, which can create example inputs to understand each model's behavior, then use those inputs to make the merged model better. This approach works after any merging method and consistently improves results without needing new training data.
What this means in practice
- •For machine learning engineers: Improve merged multi-task models without requiring access to original training data by refining merges using synthesized task-representative inputs.
- •For ai system integrators: Enhance existing model merging workflows by adding a post-merge refinement step that boosts performance across diverse tasks.
Authors
Aditya Pola, Vineeth N. Balasubramanian
Abstract
Model merging consolidates fine-tuned experts into one multi-task model without retraining. All existing data-free methods approach this problem entirely in weight space. Restricted to arithmetic on parameters, these methods never observe how each expert behaves, a signal that only emerges through forward evaluation. Accessing this behavioral signal requires inputs to evaluate on, which the data-free setting prohibits. We propose Merge++, a post-hoc method that addresses this by inverting the expert checkpoints to synthesize task-representative images, then distilling expert knowledge into the merged model using those images. Merge++ requires no additional data beyond the checkpoints themselves. It applies universally across merging algorithms and operates as a complementary refinement stage independent of the underlying weight-space method. The method consistently improves merging algorithms ranging from simple task arithmetic to state-of-the-art spectral methods, with average gains of +2 to +8 points and up to +25.9 on individual configurations.