Research map / Large language models / Agents, code and math
PREFER: Prompt Ensemble Learning via Feedback-Reflect-Refine
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes PREFER, an automatic prompt ensemble method that combines AdaBoost-style boosting driven by LLM feedback-reflect-refine prompt synthesis with bilateral (forward/backward confidence) prompt bagging, using API LLMs without any weight updates.
Problem
Existing prompt ensemble methods use a two-stage paradigm with pre-prepared prompts that require manual effort and cannot be jointly optimized for weak learners, and prompt bagging via majority voting is computationally expensive.
Approach
Transfers AdaBoost's hard-example weighting into natural-language feedback on misclassified examples, so the LLM reflects on errors and synthesizes new prompts iteratively; prompt outputs are stabilized by bilateral bagging that combines forward confidence scores with backward elimination scores in a subtractive softmax.
Key results
- PREFER achieves F1 of 0.767 on MNLI, 0.793 on QNLI and 0.744 on Liar, the best among compared methods on all seven datasets in the main table.
- Relative to APO* reproduction, PREFER improves Arsarcasm F1 from 0.639 to 0.739 and Liar from 0.658 to 0.744.
- Removing the feedback-reflect-refine module causes larger drops than removing bagging, e.g., SNLI falls from 0.647 to 0.580.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add feedback-reflect-refine prompt optimization (ablation) | PREFER · SNLI, MNLI, QNLI, RTE, Ethos, Liar, ArSarcasm | F1-score: SNLI: 0.580 -> 0.647; MNLI: 0.746 -> 0.767; QNLI: 0.720 -> 0.793; RTE: 0.659 -> 0.753; Ethos: 0.812 -> 0.963; Liar: 0.57 | Helped |
| Add bilateral bagging (ablation) | PREFER · SNLI, MNLI, QNLI, RTE, Ethos, Liar, ArSarcasm | F1-score: SNLI: 0.640 -> 0.647; MNLI: 0.713 -> 0.767; QNLI: 0.747 -> 0.793; RTE: 0.740 -> 0.753; Ethos: 0.947 -> 0.963; Liar: 0.71 | Helped |
| Replace majority voting with bilateral bagging (ablation) | PREFER · SNLI, MNLI, QNLI, RTE, Ethos, Liar, ArSarcasm | F1-score: SNLI: 0.626 -> 0.647; MNLI: 0.733 -> 0.767; QNLI: 0.767 -> 0.793; RTE: 0.760 -> 0.753; Ethos: 0.938 -> 0.963; Liar: 0.70 | Mixed |
| Use PREFER instead of the strongest reported baseline | PREFER · SNLI, MNLI, QNLI, RTE, Liar | F1-score: SNLI: 0.619 -> 0.647; MNLI: 0.746 -> 0.767; QNLI: 0.720 -> 0.793; RTE: 0.731 -> 0.753; Liar: 0.663 -> 0.744 | Helped |
| Use PREFER instead of APO | PREFER · Ethos | F1-score: 0.964 -> 0.963 | Hurt |
Surprising findings
- Removing the feedback mechanism hurts performance more than removing bilateral bagging, with majority voting being basically satisfactory.
- Adding CoT to a single prompt changes results only marginally.
Limitations
- Experiments are limited to classification and NLI tasks with 50-shot training sets.
- Relies on API LLMs (ChatGPT/GPT-4), so results depend on closed models and may not transfer to open models.
- Reported APO results partly come from the original paper and partly from the authors' reproduction, complicating comparison.
Experimental setup
Datasets: SNLI MNLI RTE QNLI Ethos Liar ArSarcasm
Benchmarks: SNLI MNLI QNLI RTE Ethos Liar ArSarcasm
Metrics: F1-score
Baselines: Single Prompt Single Prompt (CoT) Synonym Ensemble PromptBoosting APO
Builds on: AdaBoost PromptBoosting APO self-consistency In-Context Learning
Closest papers by meaning
- Unleashing the Potential of Large Language Models as Prompt Optimizers: Analogical Analysis with Gradient-based Model Optimizers (AAAI 2025 · similarity 0.88)
- GReaTer: Gradients Over Reasoning Makes Smaller Language Models Strong Prompt Optimizers (ICLR 2025 · similarity 0.88)
- Scaling Prompt Synthesis for Large Language Model Reasoning (ICML 2026 · similarity 0.87)
- Revisiting Prefix-tuning: Statistical Benefits of Reparameterization among Prompts (ICLR 2025 · similarity 0.87)
- On Speeding Up Language Model Evaluation (ICLR 2025 · similarity 0.87)
- Drag-and-Drop LLMs: Zero-Shot Prompt-to-Weights (NeurIPS 2025 · 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.