"Measuring short-form factuality in large language models"

SimpleQA tests if LLMs can answer basic facts without hallucinating

https://arxiv.org/abs/2411.04368v1

🎯 Original Problem:

Current hallucinations detection benchmarks are either too easy for modern models or too complex to evaluate accurately.

-----

🔧 Solution in this Paper:

→ SimpleQA benchmark with 4,326 short fact-seeking questions, adversarially collected against GPT-4

→ Questions designed to have single, indisputable answers that won't change over time

→ Two-stage verification process with independent AI trainers to ensure answer correctness

→ Questions cover diverse topics like science, politics, art, and technology

→ Automated grading system using ChatGPT classifier for correct, incorrect, or not attempted responses

-----

💡 Key Insights:

→ Larger models show better calibration between stated confidence and actual accuracy

→ Models consistently overstate their confidence in answers

→ Answer frequency correlates with accuracy, especially in advanced models

→ Claude models attempt fewer questions but maintain similar F-scores to GPT models

-----

📊 Results:

→ OpenAI o1-preview achieves highest score of 42.7% correct answers

→ GPT-4o scores 38.2% correct with only 1% not attempted

→ Claude-3-opus reaches 23.5% correct with 39.6% not attempted

→ Error rate estimated around 3% in benchmark questions