[論文レビュー] Prompting GPT-3 To Be Reliable

Chenglei Si, Zhe Gan+5 more|arXiv (Cornell University)|Oct 17, 2022

Topic Modeling被引用数 68

ひとこと要約

本論文は、GPT-3 の信頼性を four facets—generalizability、social biases、calibration、factuality の4つの観点で分析し、それぞれの観点を改善する prompting 戦略を開発し、より小規模な supervised モデルを上回る成果を示しています。さらにデータセット、スクリプト、予測も公開しています。

ABSTRACT

Large language models (LLMs) show impressive abilities via few-shot prompting. Commercialized APIs such as OpenAI GPT-3 further increase their use in real-world language applications. However, the crucial problem of how to improve the reliability of GPT-3 is still under-explored. While reliability is a broad and vaguely defined term, we decompose reliability into four main facets that correspond to the existing framework of ML safety and are well-recognized to be important: generalizability, social biases, calibration, and factuality. Our core contribution is to establish simple and effective prompts that improve GPT-3's reliability as it: 1) generalizes out-of-distribution, 2) balances demographic distribution and uses natural language instructions to reduce social biases, 3) calibrates output probabilities, and 4) updates the LLM's factual knowledge and reasoning chains. With appropriate prompts, GPT-3 is more reliable than smaller-scale supervised models on all these facets. We release all processed datasets, evaluation scripts, and model predictions. Our systematic empirical study not only sheds new insights on the reliability of prompting LLMs, but more importantly, our prompting strategies can help practitioners more reliably use LLMs like GPT-3.

研究の動機と目的

• Define a multi-facet reliability framework for GPT-3 aligning with ML safety concepts (generalizability, biases, calibration, factuality).
• Develop simple prompting strategies that improve GPT-3 reliability across out-of-distribution generalization, social fairness, probability calibration, and knowledge updating.
• Empirically compare GPT-3 prompting against smaller supervised models across reliability facets.
• Provide practical guidance and resources (datasets, scripts, predictions) for researchers deploying GPT-3 in real-world tasks.

提案手法

• Systematic empirical evaluation of GPT-3 prompting on four reliability facets.
• Prompt designs include random sampling of source-domain demonstrations to test generalization.
• Social bias analysis uses WinoBias and BBQ datasets with balanced prompts and natural language interventions to reduce bias.
• Calibration assessments compare LM probabilities and self-consistency with DPR-BERT baselines, using ECE and Brier scores and selective prediction.
• Knowledge updating experiments test how counterfactual prompts update memorized answers in closed-book QA and retrieval-augmented prompting.
• Multi-hop QA experiments leverage human-decomposed sub-questions to improve reasoning via Chain-of-Thought and decomposition strategies.

実験結果

リサーチクエスチョン

• RQ1 How does GPT-3 prompting affect generalization to out-of-distribution and challenging test sets compared to supervised models.
• RQ2 Can prompt design mitigate social biases and improve fairness across multiple demographic dimensions?
• RQ3 To what extent can GPT-3 calibration reflect actual accuracy, and how does prompt count influence calibration?
• RQ4 Can GPT-3 update its knowledge and reasoning chains in light of new information provided in prompts, including retrieval-augmented and decomposition approaches?
• RQ5 How can prompting strategies enhance multi-hop reasoning and factuality in QA tasks?

主な発見

• Few-shot prompting with randomly sampled demos from the source domain yields robust out-of-distribution generalization, often beating supervised baselines.
• Balanced prompts across demographic groups and natural language interventions significantly reduce bias in WinoBias and BBQ without sacrificing in-domain accuracy.
• GPT-3 demonstrates better calibration than supervised DPR-BERT when using LM probabilities or self-consistency, with selective prediction showing high accuracy on the most confident predictions.
• Counterfactual prompts enable GPT-3 to update memorized knowledge around 85% of the time, outperforming a fine-tuned T5 baseline in updating knowledge.
• Retrieval-augmented prompting (+Contriever passages) consistently improves open-domain QA accuracy across NQ, TriviaQA, and SQuAD datasets.
• Incorporating human-written decomposed sub-questions in prompts substantially improves multi-hop QA performance over standard prompting and Chain-of-Thought baselines.