Research map / Large language models / Retrieval and knowledge

ClashEval: Quantifying the tug-of-war between an LLM’s internal prior and external evidence

NeurIPS 2024 · Large language models › Retrieval and knowledge · approach: RAG robustness and evaluation · 4 citations (OpenAlex)

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 ClashEval, a 1,294-question RAG benchmark with perturbed context across six domains, measuring how six LLMs arbitrate between their priors and incorrect or correct retrieved evidence, and proposes token-probability-based correction.

Problem

RAG systems can pass incorrect retrieved content into the answer, and LLMs can cling to wrong priors when correct context is available; existing work does not measure both directions or how prior confidence and context deviation drive this choice.

Approach

Builds a benchmark where each question has both a correct and a systematically perturbed context (multipliers, year shifts, or categorical name/location changes), so both prior-bias and context-bias can be measured, then relates context preference to deviation magnitude and prior token probability and uses calibrated token-probability comparison to choose between prior and context answers.

Key results

• GPT-4o chooses the incorrect context over its correct prior over 60% of the time (context bias 0.608 when prior correct).
• Claude Opus achieves the best overall accuracy of 0.743 with context bias 0.157 and prior bias 0.021.
• Calibrated token probability correction raises GPT-4o accuracy from 0.615 to 0.754 and lowers context bias from 0.304 to 0.107, at the cost of higher prior bias (0.085).

Measured effects reported in the paper

Surprising findings

• GPT-4o exhibits higher context bias than Claude Sonnet and other smaller models despite leading general benchmarks.
• Adding more retrieved documents (k=5) lowers accuracy and reduces context bias mainly by increasing neither-chosen responses.

Limitations

• Dataset has an enriched rate of contextual errors, so bias rates are not representative of real-world rates.
• Questions are strictly fact-based and do not require multi-step reasoning or document synthesis.
• Token probability method only applies to models that expose token probabilities.

Experimental setup

Datasets: UpToDate drug information pages Wikipedia Olympic records pages Associated Press RSS news headlines Hugging Face Wikipedia 20220301.en

Benchmarks: ClashEval

Metrics: accuracy context bias prior bias context preference rate

Baselines: GPT-3.5 Llama-3-8B-Instruct Claude Sonnet Gemini 1.5 Flash random prior replacement

Builds on: LangChain RAG prompts LlamaIndex RAG prompts

Closest papers by meaning

• WikiContradict: A Benchmark for Evaluating LLMs on Real-World Knowledge Conflicts from Wikipedia (NeurIPS 2024 · similarity 0.91)
• Worse than Zero-shot? A Fact-Checking Dataset for Evaluating the Robustness of RAG Against Misleading Retrievals (NeurIPS 2025 · similarity 0.90)
• Self-RAG: Learning to Retrieve, Generate, and Critique through Self-Reflection (ICLR 2024 · similarity 0.89)
• Can LLMs Take Retrieved Information with a Grain of Salt? (NeurIPS 2026 · similarity 0.89)
• Long-Context LLMs Meet RAG: Overcoming Challenges for Long Inputs in RAG (ICLR 2025 · similarity 0.89)
• Retrieval is Not Enough: Enhancing RAG through Test-Time Critique and Optimization (NeurIPS 2025 · similarity 0.88)