🚨 New preprint!
Do LLMs really know what’s true?
In our paper,
@eliassi.bsky.social and I introduce sAwMIL: a probing method that distinguishes between true, false, and neither—capturing what LLMs actually “retain.”
We evaluated 16 open models across 3 new datasets.
📄
arxiv.org/abs/2506.23921
The Trilemma of Truth in Large Language Models
We often attribute human characteristics to large language models (LLMs) and claim that they "know" certain things. LLMs have an internal probabilistic knowledge that represents information retained d...
Jul 1, 2025 17:18
