ExposureQA: Quantifying the Factual Memory and Calibration Capabilities of Large Language Models from Pre-trained Corpora

ExposureQA is an innovative benchmark test and analysis framework focused on studying the factual recall, confidence assessment, and calibration capabilities of large language models (LLMs). Its core innovation lies in extracting "relation-aware semantic support" from pre-trained corpora, providing a new perspective for understanding how models memorize and recall facts, and aiming to address LLM factual accuracy issues (such as hallucinations, ambiguous knowledge boundaries, and mismatched confidence).

Factual Issues of Large Language Models

Large language models like GPT-4, Claude, and LLaMA perform well, but face key challenges in factual accuracy:

• Hallucination problem: Generating information that seems reasonable but is incorrect
• Ambiguous knowledge boundaries: Difficulty in determining what the model "knows" and "does not know"
• Confidence mismatch: The confidence of answers does not align with actual accuracy

Role of Pre-trained Data

LLMs' knowledge comes from massive text during the pre-training phase. Understanding how models learn, memorize, and recall facts from this data is crucial for improving model design and evaluation methods.

Relation-aware Semantic Support

The core innovation of ExposureQA is "relation-aware semantic support":

• Semantic support: Text fragments in pre-trained corpora that provide evidence or context for specific facts (e.g., sentences related to "Paris is the capital of France")
• Necessity of relation awareness: Distinguish relation types (e.g., "capital of", "located in"), consider context, and integrate multi-source evidence

Evaluation Dimensions

Evaluate LLMs from three dimensions:

1. Factual recall: Measure the accuracy, coverage, and error patterns of correct fact recall
2. Confidence: Analyze probability outputs, confidence scores, and uncertainty quantification
3. Calibration: Identify over/under confidence through calibration curves and Expected Calibration Error (ECE)

Data Construction Process

1. Corpus preprocessing: Clean and tokenize, extract fact fragments, build entity-relation indexes
2. Relation extraction: NER to locate entities, relation extraction models to identify relations, build fact triples
3. Support evidence association: Link facts to corpus positions, calculate support strength, handle multi-source support

Evaluation Methodology

• QA pair generation: Factual, reasoning, and adversarial questions
• Model evaluation protocols: Zero-shot, few-shot, and chain-of-thought evaluations

Value for Model Developers

• Diagnose model weaknesses: Identify types of poor factual performance, detect pre-trained data biases, guide data cleaning and enhancement
• Improve training strategies: Optimize sampling weights for factual data, design knowledge injection methods, improve calibration techniques

Value for Model Users

• Credibility assessment: Understand knowledge boundaries, evaluate scenario reliability, design robust prompt strategies
• Risk mitigation: Identify error sources in high-risk applications, design human-machine collaboration processes, establish output verification mechanisms

Large-scale Corpus Processing

• Challenge: TB-level data processing
• Solution: Distributed computing (Spark/Dask), memory optimization (stream processing), incremental updates

Relation Extraction Accuracy

• Challenge: Error propagation in automatic extraction
• Solution: Multi-model integration, manual verification of key samples, filtering low-confidence results

Evaluation Fairness

• Challenge: Ensure result comparability
• Solution: Standardized prompts, fixed sampling parameters, report mean and variance from multiple runs

Technical Expansion

1. Multilingual support: Evaluate cross-language factual recall
2. Temporal analysis: Track changes in factual performance across model versions
3. Domain specialization: Customization for professional fields like medicine and law

Application Deepening

• Retrieval-Augmented Generation (RAG): Evaluate factual accuracy
• Knowledge editing: Test knowledge consistency after editing
• Continual learning: Evaluate the impact of incremental learning on factual memory

ExposureQA provides a systematic framework for understanding and evaluating the factual capabilities of LLMs. By linking model performance to the semantic support from pre-trained corpora, it diagnoses the limitations of current models and points the way for designing more reliable and trustworthy AI systems. In today's era where AI is integrated into various fields of society, the assessment of factual accuracy is crucial. ExposureQA is an important attempt in this direction, and its value will be further verified and expanded in future research and applications.