We’re excited to announce the first workshop on CogInterp: Interpreting Cognition in Deep Learning Models @ NeurIPS 2025! 📣 How can we interpret the algorithms and representations underlying complex behavior in deep learning models? 🌐 coginterp.github.io/neurips2025/ 1/

New research: are prompting and activation steering just two sides of the same coin? Eric Bigelow Daniel Wurgaft Ekdeep Singh and coauthors argue they are: ICL and steering have formally equivalent effects. (1/4)

Our first Goodfire paper is out! In-context learning and activation steering are two commonly used inference-time control paradigms: we show a Bayesian Belief Update model unifies these protocols, offering a predictive theory that shows the protocols are duals of each other!

Excited to see this out! 🎉 I see this paper as an exciting step towards connecting behavior and representation of belief updating, and demonstrating the benefits of top-down accounts for interpretability (aka 'coginterp') in large-scale systems! 1/3

Great work — his is more or less how I've thought about things, but really glad to see someone formalize it and make the argument so clearly! One question I have, though, is where we should think of those latent concepts being represented across the sequence... 1/2

Very cool work led by Ekdeep Singh Lubana Can Rager Sumedh Hindupur! I'm particularly excited about how this examines some of the implicit assumptions baked into SAEs, and proposes a new approach which builds on a different foundation.