Research map / Probabilistic and scientific modeling / Bayesian optimization
Online Black-Box Prompt Optimization with Regret Guarantees under Noisy Feedback
Jinjie Fang, Runwen You, Wanli Shi, Wenkang Wang, Ganyu Wang, Haozhen Zhang, Yi Chang, Bin Gu
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 adaptive uncertainty scale adjustment on top of online zeroth-order tuning (ablation) | AOZPT · Anime / Painting, Dreamlike-2.0 | aesthetic score: 6.173 -> 6.656 (Painting), 5.861 -> 6.282 (Anime) | Improved |
| add online zeroth-order tuning alone (without scale adjustment) (ablation) | AOZPT · Anime / Painting, Dreamlike-2.0 | aesthetic score: 5.855 -> 5.861 (Anime), 6.179 -> 6.173 (Painting) | No clear effect |
| add adaptive scale adjustment (w/ delta1 vs none) (ablation) | AOZPT, Llama-3.1-8B · CNN/DailyMail | F1: 23.500 -> 25.089 -> 27.966 | Improved |
| add adaptive scale adjustment (w/ delta1 vs none) (ablation) | AOZPT, Llama-3.1-8B · GSM8K | accuracy: 69.267 -> 74.000 -> 75.533 | Improved |
| use frozen open-source LLM to generate discrete prompts instead of mapping soft prompts to vocabulary (ablation) | AOZPT · CNN/DailyMail, Llama-3.1-8B | F1: 9.911 -> 24.707 | Improved |
| use frozen open-source LLM to generate discrete prompts instead of direct vocabulary mapping (ablation) | AOZPT · CNN/DailyMail, Qwen2.5-14B | F1: 21.84 -> 24.767 | Improved |
| AOZPT versus ZO-OGD (plain online zeroth-order GD) | online prompt tuning · CNN/DailyMail and GSM8K, three LLMs | F1 / accuracy: e.g. GSM8K GPT-3.5-turbo 69.533 -> 78.133; CNN/DM Llama 24.667 -> 24.707 | Improved |
| AOZPT versus BDPL on GSM8K with Qwen2.5-14B | online prompt tuning · GSM8K, Qwen2.5-14B | accuracy: 89.000 -> 92.933 | Improved |
| AOZPT versus ZO-OGD under data drift with window L=150 | Stable Diffusion v1.5 · Anime/Painting data drift | aesthetic score: 6.037 -> 6.117 (L=150) | Improved |
| sliding-window AOZPT versus Adam, Nadam, RMSProp (ablation) | adaptive gradient update · Anime/Painting, Table 12 | aesthetic score: Painting Dreamlike: 6.179 (Adam) vs 6.140 (w=10), 6.181 (w=50) | Mixed |
| remove adaptive scale adjustment from single-point estimator variant (ablation) | AOZPT single-point · Anime, Stable Diffusion v1.5 | aesthetic score: 5.815 -> 5.872 | Improved |
| AOZPT versus MP/ICL/ZO-OGD on translation | Llama-3.1-8B · WMT14 De-En | BLEU: 30.722 -> 32.510 (vs ZO-OGD) | Improved |
Closest papers by meaning
- Posterior Inference in Latent Space for Scalable Constrained Black-box Optimization (NeurIPS 2026 · similarity 0.86)
- ROOT: Rethinking Offline Optimization as Distributional Translation via Probabilistic Bridge (NeurIPS 2025 · similarity 0.86)
- Doubly-Bounded Queue for Constrained Online Learning: Keeping Pace with Dynamics of Both Loss and Constraint (AAAI 2025 · similarity 0.86)
- No Loss, No Gain: Gated Refinement and Adaptive Compression for Prompt Optimization (NeurIPS 2025 · similarity 0.86)
- A Simple and Optimal Approach for Universal Online Learning with Gradient Variations (NeurIPS 2024 · similarity 0.86)
- Zeroth-Order Optimization Meets Human Feedback: Provable Learning via Ranking Oracles (ICLR 2024 · similarity 0.86)
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.
