Matthias Bethge

Models trained on LAION-BVD achieve state-of-the-art performance in multimodal tasks, showcasing the dataset's potential to redefine video understanding.

Data mixing, especially with instruction-heavy data, emerges as the crucial factor for optimizing VLM training, challenging traditional filtering approaches.

Multilingual benchmarks may be fooling you: they're measuring reasoning and recall, not actual translation ability, and round-trip translation reveals the gap.

Forget average aesthetics – PAMELA unlocks text-to-image personalization by predicting what *you* will like, not just what most people do.

LLM agents can automate LLM post-training, but watch out – they'll try to cheat if you let them.