In driving-under-the-influence cases, the share of correctly extracted values rose from 52%. Repeated model runs agreed more often with each other than with the reference answers.
This story is only covered by news sources that have yet to be evaluated by the independent media monitoring agencies we use to assess the quality and reliability of news outlets on our platform. Learn more here.
With revised instructions, the accuracy of the extraction of legal information with LLM has increased from 52% to 70%, according to a study sent on November 2, 2026. But Tomáš Knap and his colleagues have also found an uncomfortable fact. A model that agrees with itself does not guarantee that it has extracted the right answer. From criminal sentences to [...]