Research map / Large language models / Language, safety and interpretability
The Unreliability of Explanations in Few-shot Prompting for Textual Reasoning
Xi Ye, Greg Durrett
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 explanations before the prediction | InstructGPT · AdvHotpot | accuracy: 53.2 -> 58.2 | Improved |
| Add explanations before the prediction | InstructGPT · E-SNLI | accuracy: 56.8 -> 41.8 | Hurt |
| Use explanations in the prompt | text-davinci-002 · Synth | accuracy: 72.0 -> 86.9 | Improved |
| Add a “Let's think step by step” trigger to E-P prompts (ablation) | GPT-3 (davinci) · AdvHotpot | accuracy: 54.1 -> 50.1 | Hurt |
| Add a “Let's think step by step” trigger to E-P prompts (ablation) | InstructGPT · Synth | accuracy: 58.5 -> 58.0 | No clear effect |
| Reverse the order of the two explanation sentences (ablation) | GPT-3 (davinci) · Synth | accuracy: 49.5 -> 48.0 | Hurt |
| Reject answer-explanation pairs with nonfactual explanations | InstructGPT P-E · Synth | accuracy: 52.4% to 74.8% | Improved |
| Calibrate predictions using explanation factuality scores | InstructGPT P-E · E-SNLI | accuracy: 56.8 -> 68.5 | Improved |
| Add explanation-based features to probability calibration | InstructGPT P-E · E-SNLI (128 labels) | accuracy: 63.9 -> 68.5 | Improved |
| Calibrate answer confidence using explanation scores | InstructGPT E-P · AdvHotpot | AUC: 68.8; surpassing Few-Shot by 7 points and E-P by 4 points | Improved |
| Use factuality rather than consistency as a prediction-accuracy signal | LLM-generated explanations · Synth | accuracy/factuality agreement: 89.6 | Improved |
Closest papers by meaning
- Eliciting Causal Abilities in Large Language Models for Reasoning Tasks (AAAI 2025 · similarity 0.88)
- The Surprising Effectiveness of Test-Time Training for Few-Shot Learning (ICML 2025 · similarity 0.88)
- GIVE: Structured Reasoning of Large Language Models with Knowledge Graph Inspired Veracity Extrapolation (ICML 2025 · similarity 0.87)
- Potemkin Understanding in Large Language Models (ICML 2025 · similarity 0.87)
- In-Context Principle Learning from Mistakes (ICML 2024 · similarity 0.87)
- From Noise to Diversity: Random Embedding Injection in LLM Reasoning (NeurIPS 2026 · 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.
