AI Paper Atlas

Research map / Large language models / Language, safety and interpretability

The Unreliability of Explanations in Few-shot Prompting for Textual Reasoning

Xi Ye, Greg Durrett

NeurIPS 2022 · Large language models › Language, safety and interpretability · approach: Faithfulness of LLM explanations · 52 citations (OpenAlex)

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 changedWhereMeasuredEffect
Add explanations before the predictionInstructGPT · AdvHotpotaccuracy: 53.2 -> 58.2Improved
Add explanations before the predictionInstructGPT · E-SNLIaccuracy: 56.8 -> 41.8Hurt
Use explanations in the prompttext-davinci-002 · Synthaccuracy: 72.0 -> 86.9Improved
Add a “Let's think step by step” trigger to E-P prompts (ablation)GPT-3 (davinci) · AdvHotpotaccuracy: 54.1 -> 50.1Hurt
Add a “Let's think step by step” trigger to E-P prompts (ablation)InstructGPT · Synthaccuracy: 58.5 -> 58.0No clear effect
Reverse the order of the two explanation sentences (ablation)GPT-3 (davinci) · Synthaccuracy: 49.5 -> 48.0Hurt
Reject answer-explanation pairs with nonfactual explanationsInstructGPT P-E · Synthaccuracy: 52.4% to 74.8%Improved
Calibrate predictions using explanation factuality scoresInstructGPT P-E · E-SNLIaccuracy: 56.8 -> 68.5Improved
Add explanation-based features to probability calibrationInstructGPT P-E · E-SNLI (128 labels)accuracy: 63.9 -> 68.5Improved
Calibrate answer confidence using explanation scoresInstructGPT E-P · AdvHotpotAUC: 68.8; surpassing Few-Shot by 7 points and E-P by 4 pointsImproved
Use factuality rather than consistency as a prediction-accuracy signalLLM-generated explanations · Synthaccuracy/factuality agreement: 89.6Improved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy