Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
PromptBoosting: Black-Box Text Classification with Ten Forward Passes
Bairu Hou, Joe O'Connor, Jacob Andreas, Shiyu Chang, Yang Zhang
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 |
|---|---|---|---|
| Ensemble with AdaBoost of learned-verbalizer weak learners (vs majority vote over same learned verbalizers) (ablation) | PromptBoosting · AG's News and TREC (k=16) | accuracy (%): AG's News 83.5 -> 85.2; TREC 67.2 -> 81.6 | Improved |
| Learned verbalizer (token selection by training-set accuracy) vs random or manual verbalizer, same majority-vote ensemble (ablation) | PromptVoting · AG's News, TREC, SST-2, MR (k=16) | accuracy (%): AG's News 25.2 (random) / 67.1 (manual) -> 83.5 | Improved |
| Increase number of prompts from 1 to 10 (AdaBoost ensemble) (ablation) | PromptBoosting · TREC (k=16) | accuracy (%): 41.3 -> 81.6 | Improved |
| Increase number of prompts from 10 to 20 (ablation) | PromptBoosting · Avg. over 8 tasks (k=16) | average accuracy (%): 71.4 -> 71.3 | No clear effect |
| Smaller top-m candidate set for verbalizer search (ablation) | PromptBoosting · AG's News and RTE (k=16) | accuracy (%) and wall time: AG's News 85.2 -> 84.4 (8 -> 2 min); RTE 60.0 -> 59.1 (4 -> 0.7 min) | Mixed |
| Use PromptBoosting (black-box, 10 forward passes) vs BBTv2 (8000 forward passes) | PromptBoosting · AG's News and RTE (k=16) | wall time and accuracy: AG's News 90 min -> 8 min; RTE 70 min -> 4 min; accuracy 83.6->85.2 and 53.2->60.0 | Improved |
| Prompt set from LM-BFF, PET, or manual writing instead of the generated prompts (ablation) | PromptBoosting · AG's News and RTE (k=16) | accuracy (%): AG's News 84.7-85.4; RTE 58.2-60.6 across prompt sets | No clear effect |
| Prompt refinement by selecting top-10 prompts by validation accuracy (ablation) | PromptBoosting · SNLI, MNLI, QNLI | few-shot accuracy | Improved |
| Full training data instead of 16-shot | PromptBoosting vs fine-tuning · SST-2, MR, TREC, RTE (full data) | accuracy (%): SST-2 94.1 vs 95.5; TREC 90.5 vs 97.2; RTE 71.7 vs 81.9 (fine-tuning higher) | Hurt |
| Train on train+validation (PromptBoosting-32) vs standard k=16 split | PromptBoosting · Avg. 8 tasks (k=16) | average accuracy (%): 71.4 -> 71.9 | Improved |
Closest papers by meaning
- Language Models as Black-Box Optimizers for Vision-Language Models (CVPR 2024 · similarity 0.87)
- ProAPO: Progressively Automatic Prompt Optimization for Visual Classification (CVPR 2025 · similarity 0.86)
- Drag-and-Drop LLMs: Zero-Shot Prompt-to-Weights (NeurIPS 2025 · similarity 0.85)
- PREFER: Prompt Ensemble Learning via Feedback-Reflect-Refine (AAAI 2024 · similarity 0.85)
- SEC-Prompt:SEmantic Complementary Prompting for Few-Shot Class-Incremental Learning (CVPR 2025 · similarity 0.85)
- Localized Zeroth-Order Prompt Optimization (NeurIPS 2024 · 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.
