Large language models can 'hallucinate' factually incorrect outputs, presenting significant risks for their adoption to high-stakes applications. Jannik will present joint work recently published in Nature on detecting hallucinations in large language models using semantic entropy, which mitigates hallucinations by quantifying the model's own uncertainty over the meaning of generations. He will also discuss a recent pre-print that proposes a method to drastically reduce the cost of uncertainty quantification in LLMs by predicting semantic entropy from latent space, and he may ramble about uncertainties in LLMs more generally.

Abstract: Invisible image watermarking embeds information into image pixels in a way that remains imperceptible to the human eye but can still be retrieved even after significant image editing. In this talk, after a brief introduction to image watermarking, we’ll explore an approach designed to tackle this issue. Watermark Anything (ICLR 2025) reframes image watermarking as a segmentation problem. We’ll walk through the motivation behind this idea, how we developed and trained the model, the challenges we faced, and the final results. Duration approximately 45 minutes.