Excited this project is out! Using sparse feature circuits, we can explain and modify how LMs arrive at a behavior. In this thread, I want to highlight open directions where computational linguists can use sparse feature circuits. 🧵.

RT @OrgadHadas: We're presenting the Mechanistic Interpretability Benchmark (MIB) now! Come and chat - East 1205. Project led by @amuuuelle….

If you're at #ICML2025, chat with me, @sarahwiegreffe, Atticus, and others at our poster 11am - 1:30pm at East #1205! We're establishing a 𝗠echanistic 𝗜nterpretability 𝗕enchmark. We're planning to keep this a living benchmark; come by and share your ideas/hot takes!

RT @nikhil07prakash: How do language models track mental states of each character in a story, often referred to as Theory of Mind?. Our rec….

RT @jowenpetty: How well can LLMs understand tasks with complex sets of instructions? We investigate through the lens of RELIC: REcognizing….

RT @jsrozner: BabyLMs first constructions: new study on usage-based language acquisition in LMs w/ @LAWeissweiler, @coryshain. Simple inter….

RT @davidbau: Dear MAGA friends,. I have been worrying about STEM in the US a lot, because right now the Senate is writing new laws that cu….

RT @_joestacey_: We have a new paper up on arXiv! 🥳🪇. The paper tries to improve the robustness of closed-source LLMs fine-tuned on NLI, as….

We still have a lot to learn in editing NN representations. To edit or steer, we cannot simply choose semantically relevant representations; we must choose the ones that will have the intended impact. As @peterbhase found, these are often distinct.

By limiting steering to output features, we recover >90% of the performance of the best supervised representation-based steering methods—and at some locations, we outperform them!.

We define the notion of an “output feature”, whose role is to increase p(some token(s)). Steering these gives better results than steering “input features”, whose role is to attend to concepts in the input. We propose fast methods to sort features into these categories.

SAEs have been found to massively underperform supervised methods for steering neural networks. In new work led by @dana_arad4, we find that this problem largely disappears if you select the right features!.

RT @tomerashuach: 🚨New paper at #ACL2025 Findings!.REVS: Unlearning Sensitive Information in LMs via Rank Editing in the Vocabulary Space.….

RT @tal_haklay: Our paper "Position-Aware Circuit Discovery" got accepted to ACL! 🎉. Huge thanks to my collaborators🙏.@OrgadHadas @davidbau….

RT @boknilev: BlackboxNLP will be co-located with #EMNLP2025 in Suzhou this November! 📷This edition will feature a new shared task on circu….

RT @weGotlieb: 📣Paper Update 📣It’s bigger! It’s better! Even if the language models aren’t. 🤖New version of “Bigger is not always Better: T….

RT @babyLMchallenge: Close your books, test time!.The evaluation pipelines are out, baselines are released and the challenge is on. There i….

Presenting sparse feature circuits today at 3pm-5:30pm! Come say hi at poster #495

RT @yanaiela: 💡 New ICLR paper! 💡."On Linear Representations and Pretraining Data Frequency in Language Models":. We provide an explanation….

. Jing Huang, Rohan Gupta, @YNikankin, @OrgadHadas, @nikhil07prakash, @anja_reu, @arunasank, @ShunShao6, @alesstolfo, @mtutek, Amir Zur, @davidbau, and @boknilev!.

This was a huge collaboration with many great folks! If you get a chance, be sure to talk to Atticus Geiger, @sarahwiegreffe, @dana_arad4, @IvanArcus, @adambelfki, @yiksiux, @jadenfk23, @tal_haklay, @michaelwhanna, . .