Research Highlights

Closer looks at high-impact work built with NNsight — what the authors found, and the interventions that got them there.

Built with NNsightProtein Biology

A protein language model learned biology nobody taught it

Sparse autoencoders surface up to 2,548 interpretable features per layer inside ESM-2, and steering experiments show the model propagates them.

Nature Methods 22, 2107–2117 (2025)

Read the highlight

Built with NNsightAI Safety

Refusal isn't one direction. It's a cone.

Refusal in an aligned language model is mediated by multi-dimensional concept cones, not a single direction — so ablating one direction removes one route to refusal, not the mechanism.

ICML 2025 · PMLR vol. 267, 66945–66970

Read the highlight