Selecting Better Samples from Pre-trained LLMs: A Case Study on Question Generation

1 Introduction & Related Work

• Prompt-based learning verbalizes an input as a natural-language prompt for a pre-trained LLM in a zero-shot setting.The predicted response can then be compared with a ground-truth response for evaluation.Source: §1 · p. 1
• Large models and API-style access make domain adaptation impractical, while prompt tuning often depends on ground-truth labels that may be scarce.The introduction frames these as two distinct barriers to adapting general-domain LLMs.Source: §1 · p. 1
• The study uses question generation as a case study to reduce dependence on model access and annotations in real-world LLM use.The stated goal is to obtain better outputs from an inference-only, black-box LLM by selecting among sampled candidates without model adaptation or human annotations.Source: §1 · pp. 1–2
• Question generation conditions a natural-language question on context and an answer, with the context serving as supporting evidence.The introduction also identifies education, FAQ generation, and information seeking as applications.Source: §1 · p. 2

2 Problem Setting

• The selector chooses the sampled question maximizing an evaluation metric, with success defined as outperforming both the sampled-set mean and greedy output.The sampled set contains k questions, while greedy generation selects the most probable token at each step.Source: §2 · p. 2
• Experiments use SQuAD sentence-level question generation and Fairytale QA paragraph-level generation, with 11,877 and 1,007 test examples, respectively.Fairytale QA answers need not appear as substrings of the context; the study uses test splits because it performs no model or prompt tuning.Source: §2 · p. 2
• GPT-3 generates and selects questions in a zero-shot setup, and the study assumes access to stochastic outputs without tuning the generator.The figure caption identifies a prompt asking GPT-3 to rate question relevance.Source: §2 · p. 2, Figure 1 · p. 3
• Selections are evaluated against reference questions with BLEU-4 on SQuAD and ROUGE-L on Fairytale QA, alongside human ratings on a subset.The authors note that holistic overall ratings may be subjective and less likely to produce annotator agreement.Source: §2 · p. 2

3 Method

• Each method scores the five sampled questions and selects the candidate with the highest score.The section describes three question-selection methods.Source: §3 · p. 3
• N-gram similarity scores a question's relevance to its context using their n-grams, with n ranging from 1 to 5.The method extracts unique n-grams from both the context and question.Source: §3 · p. 3
• Round-trip selection uses a QA model to answer each generated question, then chooses the question whose recovered answer most overlaps the original answer.It measures answer overlap with F1 for SQuAD and ROUGE-L for Fairytale QA.Source: §3 · p. 3
• Prompt-based scoring asks GPT-3 to explain its judgment before choosing a rating option, a step observed to encourage output diversity.The method also defines OPS as GPT-3's overall score and APS as the average across individual meta-questions.Source: §3 · pp. 3–4
• Human evaluation results average three annotators' scores and normalize them per column across SQuAD and Fairytale QA.The displayed dimensions include seven question-quality ratings, their average, and an overall human rating.Source: Figure 2 · p. 4

4.1 Reference-based evaluation

• Round-trip outperformed greedy generation on both datasets, while every selection method beat the average score of five sampled questions.Uniform-weight ensembles with n-gram and/or prompt-based scores further improved reference-based results.Source: §4.1 · p. 4

4.2 Human Evaluation

• Human ratings show n-gram similarity performed best among individual methods on SQuAD, whereas APS led on Fairytale QA.The human evaluation comprised 16,800 annotations from 87 annotators across the two datasets.Source: §4.2 · pp. 4–5
• The authors attribute the dataset-specific trend to SQuAD’s short, simple contexts and extractive answers versus Fairytale QA’s multiparagraph contexts and abstractive, open-ended answers.They argue that local context paraphrasing suits SQuAD better, while n-gram similarity is less suitable for selecting questions about abstract ideas.Source: §4.2 · p. 5

4.3 Limitations and Future Work

• The authors caution that downstream applications may involve substantial out-of-distribution performance variance and harmful or toxic content, particularly in educational settings.Source: §4.3 · p. 5
• At the time of writing, access to GPT-3 was prohibitively restrictive; the authors identify more accessible LLMs such as GPT-J and BLOOM as future-work possibilities.Source: §4.3 · p. 5

• Using question generation as a case study, the authors propose two prompt-based approaches that assess question quality from multiple perspectives.Source: §5 · p. 5
• The approaches require neither model fine-tuning nor human annotation, and experiments with automatic and human evaluations find them effective for selecting high-quality questions.Source: §5 · p. 5

• Appendix A reports the prompt templates used in the work.Source: Contents in Appendices · p. 9
• Appendix B provides details on the human study.Source: Contents in Appendices · p. 9
• Appendix C provides the full set of experiment results.Source: Contents in Appendices · p. 9
• Appendix D reports implementation details.Source: Contents in Appendices · p. 9

• Figure 3 shows an example question-generation prompt.Source: A Prompt Designs · p. 9
• Figure 4 shows an example QA prompt used in round-trip selection.Source: A Prompt Designs · p. 9
• Figure 1 shows an example prompt for obtaining prompt scores.Source: A Prompt Designs · p. 9

B Human Study

• 50 documents per dataset were sampled, each paired with one ground-truth question and six GPT-3-generated questions: five stochastic samples and one greedy output.Three annotators rated each question on seven meta-questions and an overall rating, yielding 16,800 annotations from 87 English-speaking annotators.Source: B Human Study · p. 9
• A 15.4% spam rate was observed, and reported human scores exclude annotations removed by spam filtering.Source: B Human Study · p. 9
• Annotators rated each question using eight meta-questions, with three different annotators contributing to each meta-question rating.The reported agreement rate averages agreement among the three annotators.Source: B Human Study · p. 9

C Additional Results

• Table 3 presents reference-based evaluation scores for the question-selection methods.Source: C Additional Results · p. 9, Table 3 · p. 9
• Table 2 reports averaged agreement among three annotators, where agreement means all three chose the same option for a meta-question.Source: Table 2 · p. 9
• The paper reports that Table 4 contains the full results for human evaluation on SQuAD, while Table 5 contains the full results for human evaluation on Fairytale QA.Source: C Additional Results · p. 9

D Implementation Details

• All experiments use GPT-3 text-davinci-002, a 175B-parameter variant whose inserting-completion support facilitates question generation.Source: D Implementation Details · p. 10
• Question-generation candidates are sampled at temperature 0.7, while QA round-trip and prompt-score cases use greedy generation at temperature 0.Source: D Implementation Details · p. 10
• Human-evaluation tables report grammatical correctness, offensiveness, clarity, relevance, importance, specificity, answerability, averaged human rating, and overall human rating.Source: Table 4 · p. 12, Table 5 · p. 12