Research map / Vision-language and multimodal / CLIP and zero-shot
Unlearning without Forgetting: Securely Removing Targeted Concepts from Large-Scale Vision-Language Open-Vocabulary Detectors
Zhongze Wu, Xiu Su, Feng Yang, Dan Niu, Shan You, Yueyi Luo, Jun Long
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 |
|---|---|---|---|
| constrain updates to the null-space of retain embeddings and add one-step mean-flow forgetting on top of null-space only (ablation) | SafeDetect · UOD-Bench OD 1% forget, LLM-Det Swin-T | U-Score: 16.7 -> 22.2 | Improved |
| add cross-modal decoupling loss on decoder queries (ablation) | SafeDetect · UOD-Bench OD 1% forget, LLM-Det Swin-T | U-Score: 22.2 -> 23.5 | Improved |
| add cross-modal decoupling loss on decoder queries (ablation) | SafeDetect · UOD-Bench OD 10% forget, LLM-Det Swin-T | U-Score: 8.6 -> 9.7 | Improved |
| add null-space projection to preserve zero-shot generalization (ablation) | SafeDetect · LVIS-minival, average AP drop | LVIS AP drop: 9.6 -> 3.8 | Improved |
| apply decoupling to decoder query features rather than bbox head features (ablation) | SafeDetect · UOD-Bench OD 1% forget, LLM-Det Swin-T | U-Score: 19.8 -> 23.5 | Improved |
| decouple on bbox head features instead of no decoupling (ablation) | SafeDetect · UOD-Bench OD 1% forget, LLM-Det Swin-T | U-Score: 22.2 -> 19.8 | Hurt |
| use SafeDetect instead of NPO for unlearning | LLM-Det Swin-T · UOD-Bench OD 1% forget, forget mAP | forget mAP@50: 50.5 -> 17.8 | Improved |
| use SafeDetect instead of NPO, test with synonym class names | LLM-Det Swin-T · UOD-Bench OD 1% forget, forget mAP with synonyms | forget mAP@50: 65.8 -> 20.7 | Improved |
| use SafeDetect instead of NPO | LLM-Det Swin-T · LVIS AP after unlearning, average drop | LVIS AP drop: 14.2 -> 8.3 | Improved |
| use gradient ascent (GA) for unlearning | LLM-Det Swin-T · UOD-Bench OD 1% forget | U-Score: 58.0 -> 57.2 forget mAP; U-Score 2.7 | No clear effect |
| use SafeDetect instead of NPO for training convergence | unlearning training · UOD-Bench training loss | steps to stabilize: 750 -> 500 steps | Improved |
| use SafeDetect instead of NPO for compute cost | unlearning training · UOD-Bench | GPU-hours: 22.0 -> 13.2 | Improved |
Closest papers by meaning
- Single Image Unlearning: Efficient Machine Unlearning in Multimodal Large Language Models (NeurIPS 2024 · similarity 0.89)
- SineProject: Machine Unlearning for Stable Vision-Language Alignment (CVPR 2026 · similarity 0.88)
- From Dormant to Deleted: Tamper-Resistant Unlearning Through Weight-Space Regularization (NeurIPS 2025 · similarity 0.88)
- AUVIC: Adversarial Unlearning of Visual Concepts for Multi-modal Large Language Models (AAAI 2026 · similarity 0.88)
- Targeted Forgetting of Image Subgroups in CLIP Models (CVPR 2025 · similarity 0.87)
- Learning What to Forget: Improving LLM Unlearning via Learned Token-Level Importance (NeurIPS 2026 · 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.
