Research map / Large language models / Language, safety and interpretability
Curriculum Reinforcement Learning for Black-Box Prompt Tuning via Large Language Models
Shuai Gong, Chaoran Cui, Xiaolin Dong, Chunyun Zhang, Linwei Fan
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 imitation loss Lim (ablation) | CRL-BPT · Caltech/Flowers/UCF 16-shot | average accuracy: 77.6 -> 79.7 | Improved |
| add innovation loss Lin (ablation) | CRL-BPT · Caltech/Flowers/UCF 16-shot | average accuracy: 78.0 -> 79.7 | Improved |
| add both imitation and innovation losses (ablation) | CRL-BPT · Caltech/Flowers/UCF 16-shot | average accuracy: 76.2 -> 79.7 | Improved |
| add historical loss normalization (ablation) | CRL-BPT · Caltech/Flowers/UCF 16-shot | average accuracy: 77.4 -> 79.7 | Improved |
| add relative reward calibration (ablation) | CRL-BPT · Caltech/Flowers/UCF 16-shot | average accuracy: 77.8 -> 79.7 | Improved |
| add both HLN and RRC (ablation) | CRL-BPT · Caltech/Flowers/UCF 16-shot | average accuracy: 76.6 -> 79.7 | Improved |
| use exponential schedule instead of linear schedule (ablation) | CRL-BPT · Caltech/Flowers/UCF 16-shot | average accuracy: 78.2 -> 79.7 | Improved |
| use decaying curriculum schedule instead of constant weights (ablation) | CRL-BPT · Caltech/Flowers/UCF 16-shot | average accuracy: 76.9 -> 79.7 | Improved |
| use top-quality references (Top-20) instead of bottom-quality (Bottom-20) (ablation) | CRL-BPT · Flowers/SUN/UCF | average accuracy: 68.7 -> 70.2 | Improved |
| augment reference set with self-generated high-performing prompts (ablation) | CRL-BPT · Flowers/SUN/UCF | average accuracy: 70.2 -> 69.2 | Hurt |
| enlarge reference set from 20 to 40 prompts (ablation) | CRL-BPT · Flowers/SUN/UCF | average accuracy: 70.2 -> 69.5 | No clear effect |
| use in-context learning with references instead of imitation loss (ablation) | CRL-BPT · Flowers102 convergence | accuracy after 2,000 queries: around 72% vs 74.0% | Hurt |
Closest papers by meaning
- LLM-as-a-Tutor: Policy-Aware Prompt Adaptation for Non-Verifiable RL (NeurIPS 2026 · similarity 0.87)
- AdaCuRL: Adaptive Curriculum Reinforcement Learning with Invalid Sample Mitigation and Historical Revisiting (AAAI 2026 · similarity 0.87)
- Reward Is Enough: LLMs Are In-Context Reinforcement Learners (ICLR 2026 · similarity 0.86)
- CurES: From Gradient Analysis to Efficient Curriculum Learning for Reasoning LLMs (ICLR 2026 · similarity 0.86)
- Reward-Guided Prompt Evolving in Reinforcement Learning for LLMs (ICML 2025 · similarity 0.86)
- Scaling Prompt Synthesis for Large Language Model Reasoning (ICML 2026 · 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.
