CarveVFM
Carving Specialists Out of Generalists
Extract a compact, dense encoder by jointly considering capacity allocation, removal structure, and how well removed behavior can be recovered.
The manuscript evaluates models from DINOv3 and MetaCLIP-2 across twelve benchmarks. The visualization below isolates capacity allocation; the full method also searches operators and performs recovery.
than the uncompressed model
Method overview
Layer capacity allocation
Uniform compression gives every layer the same width. Task-aware allocation preserves more capacity in layers that matter to the task.
The paper jointly searches capacity allocation, removal operators, and recoverability. The chart isolates the allocation idea.
Method & experiments ↗Schematic allocation under a shared capacity budget. No accuracy or latency is predicted.
Related work
Contact
You can email me about my research or a possible collaboration.