🚀 Check out our draft on localizing knowledge in DiTs! Using attention contribution, we find that styles, objects, and facts can be localized to a small set of layers—though more diffusely than in UNets.

🖋️ Curious how writing differs across (research) cultures? 🚩 Tired of “cultural” evals that don't consult people? We engaged with researchers to identify & measure ✨cultural norms✨in scientific writing, and show that❗LLMs flatten them❗ 📜 arxiv.org/abs/2506.00784 1/11

This paper tests how well LLMs handle simple probability tasks. Bigger models usually do better than smaller ones. The setup gives counts for each outcome or a raw list of samples, then asks for the most common outcome, a probability table, or new samples. Questions about the