The world is being quietly rearranged by people who write very long documents.


The title they went with Generative Score Inference for Multimodal Data Noisy translates that to

New method for detecting when AI systems make up false information


Researchers developed a technique that uses AI-generated samples to measure how confident a language model or image system should be in its answers, making it easier to catch when these systems hallucinate or produce unreliable outputs. In practice: if you ask an AI system a question, this method can now tell you whether the answer is trustworthy or likely false, without requiring researchers to hand-write special rules for each new task.
Current AI systems have no built-in way to say 'I don't know' or 'I'm probably wrong here' — they confidently generate plausible-sounding false information. This method uses the same generative machinery that powers the AI itself to estimate when outputs are unreliable, potentially reducing the cost and difficulty of deploying large language models in domains where mistakes matter (medicine, law, finance).

If you insist
Read the original →