Research map / Vision-language and multimodal / Human and embodied modeling
Retrieving Versus Understanding Extractive Evidence in Few-Shot Learning
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.
Analyzes whether LLM label prediction errors are tied to failures in retrieving human-annotated extractive evidence in few-shot settings, using GPT-4 and Gemini-1.5 Pro across five datasets.
Problem
Verification of LLM document-level decisions requires that the model retrieve and correctly interpret within-document evidence; it is unclear whether prediction errors stem from evidence retrieval failures or from misinterpreting retrieved evidence.
Approach
Prompt closed LLMs to self-rationalize with extractive quotes under varying conditions (predict-then-explain, explain-then-predict, evidence given, evidence occluded) and compare quotes to human gold evidence to attribute label errors to retrieval versus interpretation failures via contingency analyses.
Key results
- Evidence quoting failure rates are low for most datasets (e.g., 0.33% to 4.33%) but reach 40.33% for GPT-4-turbo on Evidence Inference.
- Self-rationalization has marginal effects on label accuracy for most datasets, with the largest gains on Evidence Inference (+7.9 points for GPT-4) and HealthFC.
- Label errors are mostly associated with high evidence recall (misinterpretation or confounding evidence) rather than missing key evidence in MultiRC, SciFact, and WikiAttack.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add self-rationalization by explaining before predicting | GPT-4 and Gemini-1.5 · Evidence Inference and HealthFC | label accuracy: +7.9/5.0 for GPT-4 and +6.4/3.2 for Gemini-1.5 | Helped |
| Add self-rationalization by explaining before predicting | GPT-4 and Gemini-1.5 · MultiRC, SciFact, and WikiAttack | label accuracy | Mixed |
| Change operation order from predict-then-explain to explain-then-predict (ablation) | GPT-4 · MultiRC, SciFact, and HealthFC | label accuracy | Mixed |
| Change operation order from predict-then-explain to explain-then-predict (ablation) | Gemini-1.5 · All five datasets | label accuracy | No clear effect |
| Provide human-annotated evidence in the prompt (ablation) | GPT-4 and Gemini-1.5 · SciFact | label accuracy: 0.843 -> 0.947; 0.820 -> 0.947 | Helped |
| Provide human-annotated evidence in the prompt (ablation) | GPT-4 and Gemini-1.5 · HealthFC | label accuracy: 0.733 -> 0.947; 0.787 -> 0.857 | Helped |
| Provide human-annotated evidence in the prompt (ablation) | GPT-4 and Gemini-1.5 · MultiRC | label accuracy: 0.893 -> 0.887; 0.892 -> 0.885 | No clear effect |
| Provide human-annotated evidence without the surrounding document (ablation) | GPT-4 and Gemini-1.5 · WikiAttack | label accuracy: 0.777 -> 0.917; 0.703 -> 0.901 | Helped |
| Remove human-annotated evidence from the document (ablation) | GPT-4 and Gemini-1.5 · MultiRC | label accuracy: 0.893 -> 0.663; 0.892 -> 0.666 | Hurt |
| Remove human-annotated evidence from the document (ablation) | GPT-4 and Gemini-1.5 · SciFact | label accuracy: 0.843 -> 0.633; 0.820 -> 0.587 | Hurt |
| Remove human-annotated evidence from the document (ablation) | GPT-4 and Gemini-1.5 · WikiAttack | label accuracy: 0.777 -> 0.533; 0.703 -> 0.563 | Hurt |
| Remove human-annotated evidence from the document (ablation) | GPT-4 and Gemini-1.5 · Evidence Inference | label accuracy: 0.809 -> 0.721; 0.767 -> 0.698 | Hurt |
Surprising findings
- Evidence retrieval error is mostly not associated with evidence interpretation error, with models typically able to interpret human evidence when it is provided.
- Forcing explanation before prediction slightly reduced label accuracy for GPT-4 on several datasets, while prediction-then-explanation was mostly neutral or better.
Limitations
- Analysis uses random 300-item subsets of development sets to limit API cost.
- Only two closed proprietary models were evaluated; no open models.
- Relabeling of HealthFC was performed manually by the authors, and several contingency comparisons are small in count.
Experimental setup
Datasets: MultiRC SciFact WikiAttack Evidence Inference HealthFC
Benchmarks: MultiRC SciFact WikiAttack Evidence Inference HealthFC
Metrics: label accuracy evidence F1 evidence recall evidence precision quote failure rate
Baselines: zero-shot prompting few-shot prompting
Builds on: chain-of-thought ERASER
Closest papers by meaning
- ClashEval: Quantifying the tug-of-war between an LLM’s internal prior and external evidence (NeurIPS 2024 · similarity 0.86)
- Worse than Zero-shot? A Fact-Checking Dataset for Evaluating the Robustness of RAG Against Misleading Retrievals (NeurIPS 2025 · similarity 0.86)
- Copyright Traps for Large Language Models (ICML 2024 · similarity 0.85)
- Learning Evidence Highlighting for Frozen LLMs (NeurIPS 2026 · similarity 0.85)
- BayesRAG: Probabilistic Mutual Evidence Corroboration for Multimodal Retrieval-Augmented Generation (NeurIPS 2026 · similarity 0.85)
- Distinguishing the Knowable from the Unknowable with Language Models (ICML 2024 · similarity 0.84)
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.