Research map / Vision-language and multimodal / CLIP and zero-shot
Branch, or Layer? Zeroth-Order Optimization for Continual Learning of Vision-Language Models
Ziwei Liu, Borui Kang, Wei Li, Hangjie Yuan, Yanbing Yang, Wenbin Li, Yifan Zhu, Tao Feng, Jun Luo
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 |
|---|---|---|---|
| replace FO with ZO in both branches (ablation) | MoE4Adapter VLCL · CIFAR-100 Inc20 | Last accuracy: 80.47 -> 69.29 | Hurt |
| apply ZO only in language branch instead of both (ablation) | MoE4Adapter VLCL · CIFAR-100 Inc20 | Last accuracy: 69.29 -> 80.94 | Improved |
| apply ZO in language branch vs vision branch (ablation) | MoE4Adapter VLCL · TinyImageNet Inc20 | Last accuracy: 49.65 -> 49.14 | Mixed |
| layer-wise interleaved ZO (Hop-odd) in dual branch instead of all-layer ZO (ablation) | MoE4Adapter VLCL · CIFAR-100 Inc20 | Last accuracy: 69.29 -> 81.64 | Improved |
| interleaved layer ZO (Hop-even) in vision branch vs full-FO baseline (ablation) | MoE4Adapter VLCL · CIFAR-100 Inc20 | Last accuracy: 80.47 -> 82.41 | Improved |
| contiguous Prefix (six) ZO vs interleaved Hop-odd in dual branch (ablation) | MoE4Adapter VLCL · CIFAR-100 Inc20 | Last accuracy: 81.64 -> 79.12 | Hurt |
| add MoZO (sign gradient + smaller vision perturbation) to Hop-even dual (ablation) | Dual w/ Hop-even · TinyImageNet Inc20 | Avg accuracy: 60.53 -> 61.23 | Improved |
| add MoZO to Hop-odd dual on CIFAR Inc10 (ablation) | Dual w/ Hop-odd · CIFAR-100 Inc10 | Avg accuracy: 86.98 -> 87.02 | No clear effect |
| add loss-validated conservative ZO instead of naive single-estimate ZO (ablation) | Language w/ ZO · CIFAR-100 Inc10 | Last accuracy: 76.19 -> 76.74 | Improved |
| add gradient sign transform to naive ZO in language branch (ablation) | Language w/ ZO* · CIFAR-100 Inc10 | Last accuracy: 76.19 -> 78.52 | Improved |
| apply ZO instead of FO in dual branch (memory) | MoE4Adapter VLCL · GPU memory | memory: ~19.96GB -> ~2.17GB | Improved |
Closest papers by meaning
- SharpZO: Hybrid Sharpness-Aware Vision Language Model Prompt Tuning via Forward-Only Passes (NeurIPS 2025 · similarity 0.89)
- Stabilizing Zero-Shot Prediction: A Novel Antidote to Forgetting in Continual Vision-Language Tasks (NeurIPS 2024 · similarity 0.88)
- Synthetic Data is an Elegant GIFT for Continual Vision-Language Models (CVPR 2025 · similarity 0.88)
- ZIP: An Efficient Zeroth-order Prompt Tuning for Black-box Vision-Language Models (ICLR 2025 · similarity 0.87)
- Frustratingly Easy Test-Time Adaptation of Vision-Language Models (NeurIPS 2024 · similarity 0.87)
- Advancing Cross-domain Discriminability in Continual Learning of Vision-Language Models (NeurIPS 2024 · similarity 0.87)
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.
