Tried steering with SAEs and found that not all features behave as expected? Check out our new preprint - "SAEs Are Good for Steering - If You Select the Right Features" 🧵

1/5 🚨 New paper alert! StressTest: Can YOUR Speech LM Handle the Stress? Sentence stress = emphasis on words to signal intent, contrast, or new info. We built StressTest — a benchmark for testing stress reasoning.🗣️💬 Then, meet StresSLM who finally gets it! Insights & Links 👇

When we "erase" a concept from a diffusion model, is that knowledge truly gone? 🤔 We investigated, and the answer is often 'no'! Using simple probing techniques, the knowledge traces of the erased concept can be easily resurfaced 🔍 Here is what we learned 🧵👇

🇵🇹 Spoke at #DH2025 about Religious Journeys in Holocaust Testimonies (arXiv link in thread) 🐟 Connecting with researchers using novel computational tools on real-world challenges in the humanities was inspiring! 🏰 Excited to build on these interdisciplinary methods!

Presenting my poster : 🕊️ DOVE - A large-scale multi-dimensional predictions dataset towards meaningful LLM evaluation, Monday 18:00 Vienna, #ACL2025 Come chat about LLM evaluation, prompt sensitivity, and our 250M COLLECTION OF MODEL OUTPUTS!

🚨 Benchmarks tell us which model is better — but not why it fails. For developers, this means tedious, manual error analysis. We're bridging that gap. Meet CLEAR: an open-source tool for actionable error analysis of LLMs. 🧵👇