🚨🚨 New preprint 🚨🚨 Ever wonder whether CoTs correspond to the internal reasoning process of the model? We propose a novel parametric faithfulness approach, which erases information contained in CoT steps from parameters to assess CoT faithfulness. arxiv.org/abs/2502.14829

1/14 🎉 Excited to announce that our paper, "DEPTH: Discourse Education through Pre-Training Hierarchically", has been accepted to #Rep4NLP at #NAACL2025!!! Joint work with Ofek Glick , Chaim Baskin and Yonatan Belinkov

🚨 It's often claimed that LLMs know more facts than they show in their outputs, but what does this actually mean, and how can we measure this “hidden knowledge”? In our new paper, we clearly define this concept and design controlled experiments to test it. 1/🧵

🎉 Our Actionable Interpretability workshop has been accepted to #ICML2025! 🎉 >> Follow Actionable Interpretability Workshop ICML2025 Tal Haklay Anja Reusch Marius Mosbach Sarah Wiegreffe Ian Tenney (@[email protected]) Mor Geva Paper submission deadline: May 9th!

Check out our new work on how information flows in text-to-image models! Turns out, the text encoder isn’t doing what you’d expect — and that has real consequences for model performance and errors. For a deeper dive, see Guy Kaplan’s post. Paper link is in the first comment!

Very happy to have been a part of this effort to standardize eval in mechanistic interpretability! Plenty of resources in the thread & a lot of collaborators at ICLR/NAACL (they won't be wearing black, though)

Slides available here: docs.google.com/presentation/d…

BlackboxNLP will be co-located with #EMNLP2025 in Suzhou this November! 📷This edition will feature a new shared task on circuits/causal variable localization in LMs, details: blackboxnlp.github.io/2025/task If you're into mech interp and care about evaluation, please submit!

BlackboxNLP, the leading workshop on interpretability and analysis of language models, will be co-located with EMNLP 2025 in Suzhou this November! 📆 This edition will feature a new shared task on circuits/causal variable localization in LMs, details: blackboxnlp.github.io/2025/task

🚨New paper at #ACL2025 Findings! REVS: Unlearning Sensitive Information in LMs via Rank Editing in the Vocabulary Space. LMs memorize and leak sensitive data—emails, SSNs, URLs from their training. We propose a surgical method to unlearn it. 🧵👇w/Yonatan Belinkov Martin Tutek 1/8

Really cool work introducing a gradient-free method for unlearning organically memorized sensitive information from LMs! (we also curate two datasets of organically memorized sensitive information) Check out the 🧵 below and come talk to us at ACL 2025 in Vienna 🍻

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" 🧵

Delighted that ✨Mor Geva (Mor Geva) and ✨Anna Ivanova (Anna Ivanova) will complete our speaker lineup and talk about the INTERPLAY of model internals and behavior. Be there and submit by June 30th 📄 shorturl.at/sBomu See you in 🇨🇦 Conference on Language Modeling #nlproc #interpretability

🚨 Ever wondered how much you can ace popular MCQ benchmarks without even looking at the questions? 🤯 Turns out, you can often get significant accuracy just from the choices alone. This is true even on recent benchmarks with 10 choices (like MMLU-Pro) and their vision