Research map / Architectures and efficiency / RNNs and normalization
Revisiting Model Stitching In the Foundation Model Era
Zheda Mai, Ke Zhang, Fu-En Wang, Zixiao Ken Wang, Albert Y. C. Chen, Lu Xia, Min Sun, Wei-Lun Chao, Cheng-Hao Kuo
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| pre-train stitch layer with final feature matching before task loss training (ablation) | DINOv2->SigLIP2 MLP stitch · fMoW, stitch layer 2 | accuracy (%): 25.1 -> 51.7 | Improved |
| pre-train stitch layer with final feature matching before task loss training (ablation) | SigLIP2->DINOv2 MLP stitch · fMoW, stitch layer 2 | accuracy (%): 38.7 -> 53.8 | Improved |
| match final features (FFM) instead of layer features (LFM) at stitch position for initialization (ablation) | DINOv2->SigLIP2 · fMoW, stitch layer 2 | accuracy (%): 33.1 -> 51.7 | Improved |
| FFM vs LFM initialization at deep stitch positions (ablation) | DINOv2->SigLIP2 and SigLIP2->DINOv2 · fMoW, layers 14-22 | accuracy (%): 70.3 (LFM) vs 71.8 (FFM) at layer 22 | No clear effect |
| use stitch layer trained on layer feature matching only (low layer distance) (ablation) | SigLIP->DINOv2 · final feature distance | L2 feature distance | Hurt |
| stitch heterogeneous VFMs instead of self-stitching the source or target (ablation) | DINOv2/SigLIP2 stitched models · fMoW stitch layer 22 | accuracy (%): 68.9 -> 71.8 | Improved |
| stitch heterogeneous VFMs instead of self-stitch (ablation) | DINOv2/SigLIP2 on ADE20K · ADE20K, layers 14 and 22 | mIoU (%): +0.5 to +0.7 mIoU | Improved |
| use CLIP as the source model in stitching | stitched VFMs with CLIP source · fMoW, DINOv2/DINOv3/SigLIP2/CLIP | accuracy (%) | Hurt |
| use MLP stitch layer instead of linear layer (ablation) | DINOv2->SigLIP2 stitching · fMoW, layer 18 | accuracy (%): 69.1 -> 72.0 | Improved |
| use LoRA stitch layer instead of MLP (ablation) | DINOv2->SigLIP2 stitching · fMoW, layer 18 | accuracy (%): 72.0 -> 67.7 | Hurt |
| train stitch layer on task-agnostic LLaVA-1.5 images instead of task-specific data (ablation) | FFM stitching, layer 22 · fMoW, iNaturalist linear probing | accuracy (%): fMoW 54.7 -> 53.8; iNat 75.5 -> 77.7 (S to D) | No clear effect |
| share early layers across VFMs (VFM Stitch Tree at layer 14) | MoF-LLaVA (CLIP + DINOv2) with Qwen 3B · VQAv2, MME | share of two-VFM gain recovered: 84% of gain at 39% extra resources | Improved |
Closest papers by meaning
- Efficient Stitchable Task Adaptation (CVPR 2024 · similarity 0.87)
- AVA-Bench: Atomic Visual Ability Benchmark for Vision Foundation Models (CVPR 2026 · similarity 0.86)
- Towards a learning theory of representation alignment (ICLR 2025 · similarity 0.86)
- Towards Robust Image Stitching: An Adaptive Resistance Learning against Compatible Attacks (AAAI 2024 · similarity 0.86)
- Functional Alignment Can Mislead: Examining Model Stitching (ICML 2025 · similarity 0.85)
- Rethinking Model Selection in VLM Through the Lens of Gromov-Wasserstein Distance (CVPR 2026 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
