LLM entity extraction unreliable, study finds

A new arXiv paper investigates the reliability of large language models (LLMs) in extracting information, specifically from cyber threat intelligence documents. The study found that even advanced models like GPT-6.1-Sol, DeepSeek-V4 Flash, and GLM-5.3-Flash exhibit significant unreliability, with low precision and high calibration errors. Contrary to common assumptions, the primary failure mode is not hallucination but rather background mentions and incorrect scope, with actual unsupported candidates being rare. While confidence signals can help prioritize review, they do not guarantee accuracy, as even minor sentence perturbations can drastically alter high-confidence judgments. AI

IMPACT Highlights limitations in LLM data extraction, suggesting caution for high-stakes applications like threat intelligence.

RANK_REASON Academic paper detailing LLM evaluation and failure modes. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• arXiv
• CatalyzeX
• Connected Papers
• cyber threat intelligence
• DagsHub
• DeepSeek-V4 Flash
• Gotit.pub
• Hugging Face
• Litmaps
• ScienceCast
• Scite

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →