Goodfire's Neural Geometry series: manifold steering and an adder module inside Llama 3.1 8B
burny_tech · x · 2026-11-10
Goodfire published its Neural Geometry Series, arguing that the multidimensional geometric structures in neural activations are a crucial frontier for understanding and controlling AI.
- Activations mirror world structure: days of the week form circles, colors an HSL manifold, the tree of life appears in genomic representations.
- Steering along curved manifolds yields cleaner, more targeted behavior changes than linear steering vectors.
- They found a general-purpose addition module inside Llama 3.1 8B that manipulates circular number representations.
- Further posts study whether SAE features (straight lines) can reconstruct curved geometry via an unsupervised manifold-discovery pipeline, and how LLM representations of stories evolve temporally over a prompt.
Related event: Goodfire Launches Neural Geometry Series to Map and Control Model Internals(2 posts)→
More from Research
- Hazard-free Boolean circuits can be exponentially more expensive — and NP-hard to build — thomasahle · 2026-11-10
- Constructing hazard-free Boolean circuits proven NP-hard — thomasahle · 2026-11-10
- Unitree Go2 robot asks clarifying questions via Gemini and Bayesian active learning — DynamicWebPaige · 2026-11-10
- Visual Token Routing cuts computer-use agent inference cost up to 3.3x with 2.5-5.8x fewer tokens — wellecks · 2026-11-10
- SimFoundry auto-builds interactive 3D sim scenes from video for robot policy training — DynamicWebPaige · 2026-11-10
- CMU paper: one unmodified computer-use agent wins prizes across four embodied AI contests — wellecks · 2026-11-10