ExposureQA: Revealing LLM Fact Recall Mechanisms via Relation-Aware Analysis of Pre-trained Corpora

ExposureQA is a benchmark and analysis framework for studying fact recall, confidence, and calibration capabilities of large language models (LLMs). By extracting relation-aware semantic support from pre-trained corpora, it deeply understands the source of model knowledge and the causes of hallucinations, helping shift from 'black-box testing' to 'white-box analysis' and providing key tools for LLM optimization, evaluation, and hallucination research.

LLMs often exhibit impressive 'knowledge', but their sources are unclear and they are prone to hallucinations. Traditional evaluations only focus on answer correctness, ignoring internal knowledge representation and confidence calibration mechanisms. ExposureQA raises a core question: Can we understand the model's fact recall ability, confidence performance, and the root causes of hallucinations by tracing relevant information in pre-trained corpora? This framework is designed precisely to answer this question.

The innovation of ExposureQA lies in the concept of 'relation-aware semantic support', with steps including:

1. Entity relation extraction: Identify core entities and their relations in the question (e.g., the capital relation between Paris and France);
2. Pre-trained corpus tracing: Locate text fragments containing these relations;
3. Support quantification: Calculate exposure levels (frequency, context diversity, expression variations, etc.);
4. Correlation analysis: Correlate support levels with model answer accuracy, confidence, and calibration. Framework components include:

• Dataset: QA pairs covering various entity relations, with pre-trained support annotations and cases of facts, reasoning, and hallucinations;
• Experimental code: Scripts for pre-trained corpus retrieval, model inference, confidence extraction, calibration evaluation, and visualization;
• Metrics: Fact recall rate, confidence calibration, hallucination detection, relation-aware support level.

Through ExposureQA analysis, the following key findings are obtained:

1. Traceable knowledge sources: The model's mastery of facts is highly correlated with the frequency and diversity of exposure to those facts during pre-training;
2. Mismatch between confidence and accuracy: LLM confidence often deviates from actual accuracy; for some relation types, even if they appear frequently, the model may still be overconfident;
3. Pre-trained roots of hallucinations: Hallucination cases correspond to relation types with low support or ambiguous expressions in pre-trained corpora;
4. Differences in relation types: Distributions of relations like geography, time, and causality in pre-trained corpora differ, directly affecting model performance.

ExposureQA has important value in multiple scenarios:

• Model developers: Optimize data filtering and training strategies to improve factual accuracy;
• Model evaluators: Identify model knowledge gaps and provide fine-grained analysis tools;
• Hallucination research: Understand hallucination mechanisms and provide a basis for mitigation strategies;
• RAG system developers: Determine when to introduce external knowledge sources and select retrieval strategies.

ExposureQA has limitations:

• Restricted access to pre-trained corpora (complete data of large models is not public);
• Possible errors in entity relation extraction;
• The current version mainly focuses on English corpora and QA. Future directions: Expand to more languages, integrate more refined corpus attribution methods, and combine with knowledge editing technologies.

The core points of ExposureQA include:

1. The model's fact recall ability is closely related to the exposure level in pre-trained corpora;
2. Pre-trained distributions of different entity relation types affect model performance;
3. LLM confidence often does not accurately reflect the actual level of knowledge mastery;
4. Hallucinations can be traced to their roots through pre-trained corpus analysis;
5. Understanding LLMs requires shifting from input-output testing to white-box analysis of internal mechanisms.