Audits of large language model (LLM) entity extraction from full-length documents have repeatedly shown severe unreliability, but two questions have remained open: what the failures actually consist of, and whether confidence signals can salvage the output. A new paper on arXiv addresses both questions in the context of cyber threat intelligence.

The authors decompose the failure modes of LLM extraction, aiming to move beyond a simple accuracy score and understand the specific ways in which extraction goes wrong. They also investigate whether confidence signals—model-reported measures of certainty—can be used to identify or filter unreliable extractions.

The work is motivated by the need for usable extraction in cyber threat intelligence, where errors can have operational consequences. By breaking down failures and testing confidence signals, the study seeks to turn unreliable LLM output into something that can be trusted in practice.