Research map / Large language models / Retrieval and knowledge
Worse than Zero-shot? A Fact-Checking Dataset for Evaluating the Robustness of RAG Against Misleading Retrievals
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.
Introduces RAGuard, a political fact-checking benchmark whose retrieval corpus is naturally occurring Reddit posts labeled as supporting, misleading, or unrelated via GPT-4 behavioral annotation, and shows that LLM RAG systems underperform their zero-context baselines.
Problem
RAG benchmarks evaluate under clean or synthetically noised retrieval, overestimating robustness; LLMs struggle to reason over naturally misleading or conflicting evidence in political fact-checking.
Approach
Build a fact-checking retrieval corpus from Reddit posts about PolitiFact claims and label each document by whether a GPT-4 classifier given only that document is pushed to a correct or incorrect verdict (supporting/misleading/unrelated), then evaluate RAG settings including oracle retrieval of misleading documents.
Key results
- All tested LLM RAG systems score below their zero-context baselines; e.g., GPT-4o drops 31.7% and Mistral 58.0% under Oracle Retrieval (Misleading), with every model below 50% accuracy.
- Llama 3 falls from 62.50% zero-context to 36.81% under Oracle Retrieval (Misleading).
- Human experts and normal annotators on a 64-instance subset are reported to handle misleading context better than LLMs, with experts showing consistently high accuracy.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add top-1 retrieved context | RAGuard fact-checking across tested LLMs · RAG-1 vs Zero-Context Prediction | accuracy: 56.87 -> 52.68; 62.50 -> 59.40; 63.97 -> 59.14; 61.06 -> 56.68; 67.33 -> 64.80; 74.51 -> 70.09; 69.98 -> 66.88; 63.67 -> | Hurt |
| Add top-5 retrieved context | RAGuard fact-checking across tested LLMs · RAG-5 vs Zero-Context Prediction | accuracy: 56.87 -> 49.74; 62.50 -> 61.37; 63.97 -> 58.91; 61.06 -> 57.59; 67.33 -> 65.90; 74.51 -> 68.58; 69.98 -> 57.81; 63.67 -> | Mixed |
| Supply associated documents across all evidence categories | RAGuard fact-checking across tested LLMs · Oracle Retrieval (All) vs Zero-Context Prediction | accuracy: 56.87 -> 53.89; 62.50 -> 61.09; 63.97 -> 51.55; 61.06 -> 52.38; 67.33 -> 53.22; 74.51 -> 52.56; 69.98 -> 50.06; 63.67 -> | Hurt |
| Supply only misleading associated documents | RAGuard fact-checking across tested LLMs · Oracle Retrieval (Misleading) vs Zero-Context Prediction | accuracy: 56.87 -> 44.04; 62.50 -> 36.81; 63.97 -> 26.88; 61.06 -> 30.57; 67.33 -> 45.97; 74.51 -> 35.98; 69.98 -> 38.25; 63.67 -> | Hurt |
| Use CRAG instead of zero-shot prediction | Llama 2 · RAGuard — Oracle Retrieval | accuracy: 50.57% -> 37.24% | Hurt |
| Retrieve five documents rather than one | Standard RAG, particularly Claude 3.5 and DeepSeek · RAGuard — RAG-5 vs RAG-1 | accuracy: 70.09 -> 68.58; 66.88 -> 57.81 | Mixed |
| Increase retrieval from top 1 to top 5 | Standard RAG retrieval · RAGuard — Misleading Retrieval Recall | Misleading Retrieval Recall: 21.3% -> 44.8% | Hurt |
| Use GPT-4o rather than Claude 3.5 on misleading-only retrieval | RAGuard Oracle Retrieval (Misleading) · Oracle Retrieval (Misleading) | accuracy: 35.98 -> 45.97 | Helped |
Surprising findings
- Zero-context prediction outperforms all RAG settings for every model tested, including reasoning models.
- Corrective RAG performs worse than zero-shot Llama 2 (37.24% vs 50.57%) in the Oracle Retrieval setting, as its web-search fallback adds noise.
Limitations
- Document labels are produced by GPT-4 behavior, introducing model-specific bias; cross-model agreement only checked on a subset (Cohen's kappa 0.789 Claude, 0.650 Gemini).
- Human study is small (4 annotators, 64 instances), limiting human-model comparison strength.
- Task relies on gold PolitiFact verdicts and the dataset's misleadingness is defined relative to LLM behavior rather than humans.
Experimental setup
Datasets: PolitiFact claims Reddit posts retrieved via Google Search
Benchmarks: RAGuard
Metrics: accuracy misleading retrieval recall retrieval precision recall NDCG
Baselines: Corrective RAG Zero-Context Prediction Standard RAG (RAG-1, RAG-5) Oracle Retrieval
Builds on: Corrective RAG
Closest papers by meaning
- ClashEval: Quantifying the tug-of-war between an LLM’s internal prior and external evidence (NeurIPS 2024 · similarity 0.90)
- RAGGED: Towards Informed Design of Scalable and Stable RAG Systems (ICML 2025 · similarity 0.90)
- Retrieval is Not Enough: Enhancing RAG through Test-Time Critique and Optimization (NeurIPS 2025 · similarity 0.90)
- Self-RAG: Learning to Retrieve, Generate, and Critique through Self-Reflection (ICLR 2024 · similarity 0.89)
- Speculative RAG: Enhancing Retrieval Augmented Generation through Drafting (ICLR 2025 · similarity 0.89)
- Benchmarking Large Language Models in Retrieval-Augmented Generation (AAAI 2024 · similarity 0.88)
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.