Research map / Probabilistic and scientific modeling / Bayesian optimization
Prime Once, then Reprogram Locally: An Efficient Alternative to Black-Box Service Model Adaptation
Yunbei Zhang, Chengyi Cai, Feng Liu, Jihun Hamm
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 |
|---|---|---|---|
| add priming stage before local visual reprogramming (ablation) | AReS local ViT-B/16 · EuroSAT 16-shot | accuracy: 70.6 -> 85.7 | Improved |
| priming alone without visual reprogramming (ablation) | local ViT-B/16 · EuroSAT 16-shot | accuracy: 47.9 -> 45.6 | Hurt |
| add priming on top of local VR plus linear probing (ablation) | local ViT-B/16 · EuroSAT 16-shot | accuracy: 80.1 -> 85.7 | Improved |
| replace ZOO reprogramming of the service API with AReS on GPT-4o | GPT-4o · EuroSAT 16-shot | accuracy: 59.4 -> 87.2 | Improved |
| use ZOO-based BlackVIP on GPT-4o | GPT-4o · EuroSAT 16-shot | accuracy: 59.4 -> 60.1 | No clear effect |
| use ZOO-based BAR or BlackVIP on LLaVA | LLaVA · EuroSAT 16-shot | accuracy: 40.1 -> 34.1 (BAR), 39.4 (BlackVIP) | Hurt |
| add AReS-MS model selection between local model and zero-shot API | AReS · 10 datasets, 16-shot, CLIP ViT-B/16 | accuracy (avg): 65.4 -> 69.3 | Improved |
| replace BlackVIP with AReS | CLIP ViT-B/16 service · 10 datasets, 16-shot | API calls (M): 754.20 -> 0.02 | Improved |
| apply input-level reprogramming on Food101 | AReS · Food101 16-shot | accuracy: 85.9 -> 68.8 | Hurt |
| add priming on top of direct local visual reprogramming (ablation) | AReS ViT-B/32 local · full-shot, VM service | accuracy (avg): 45.3 -> 50.4 | Improved |
| use fewer than 8 shots for priming and local VR (ablation) | AReS · few-shot VLM EuroSAT | accuracy | Mixed |
| increase unlabeled downstream data for priming (ablation) | AReS · EuroSAT | accuracy | Improved |
Closest papers by meaning
- MESS+: Dynamically Learned Inference-Time LLM Routing in Model Zoos with Service Level Guarantees (NeurIPS 2025 · similarity 0.85)
- AugServe: Adaptive Request Scheduling for Augmented Large Language Model Inference Serving (ICML 2026 · similarity 0.85)
- Fast Inference for Augmented Large Language Models (NeurIPS 2025 · similarity 0.84)
- When Replanning Becomes the Bottleneck: Budgeted Replanning for Embodied Agents (ICML 2026 · similarity 0.83)
- Ravan: Multi-Head Low-Rank Adaptation for Federated Fine-Tuning (NeurIPS 2025 · similarity 0.83)
- Enabling VLA Action Self-Verification via VLM Token Probability Bucketing (NeurIPS 2026 · similarity 0.83)
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.
