Really enjoyed Goodfire’s neural geometry posts. A few thoughts. The current examples mostly focus on explicit variables, e.g., days of the week, and show how their geometry in activation space mirrors the geometry of the model’s outputs/behavior. (1/3)
What’s especially interesting is going beyond the final output variable, e.g., days of the week: asking what latent variables the model uses internally, how they are represented geometrically, and whether manifold-aware steering can causally test how they give rise to downstream behavior. (2/3)
Beyond that, the bigger challenge is unsupervised discovery of those latent manifolds from activations + behavior: discovering the underlying representations that drive cognitive behavior without assuming labels for those latents in advance. (3/3)
New paper 🚨 #ICLR26
Most world models predict the future from a past trajectory. But neuroscience suggests that such inference can instead be made from temporally independent experiences.
We built the Episodic Spatial World Model (ESWM), a model that does exactly this:
Video abstract [1/2]
We've updated the preprint of our Naturalistic Computational Cognitive Science paper (arxiv.org/abs/2502.20349) — we've tried to clarify and streamline the arguments, and added some new examples: 1/5
Some incredibly exciting news that still feels a bit surreal: I am starting my research group @mpicybernetics.bsky.social!
The best part? We are hiring at all levels!
Curious about the science we plan to tackle, read along :)
#NeuroSky
🚀 PhD position in #NeuroAI & neurodevelopment 🚀
Co-supervised by Sarah Lippé and myself, to investigate visual processing & cognition abnormalities in children with neurodevelopmental disorders in a neuroAI framework.
Full project details and how to apply here: tinyurl.com/kbuyntpn
🧠🤖 📈
Want to match neural representations from different days and get more trials for analyses? Interested in multi-scale neural dynamics in decision variability?
Visit our #cosyne2026 poster today afternoon (Sat)!
3-161: Dynamics-based alignment across sessions reveals latent neural computation
Somewhat late but thrilled to share that our paper on Algorithmic Primitives & Compositional Geometry of Reasoning in Language Models was accepted to ICML 2026. Huge thanks to incredible coauthors, esp @sflippl.bsky.social @eberleoliver.bsky.social @thomasmcgee.bsky.social.
Camera ready coming soon.