Search papers, labs, and topics across Lattice.
Affiliation:
3
0
5
Parameter matrices are no longer inscrutable black boxes: decomposing weights directly into the activation features they read and write enables scalable, causally editable circuit discovery at 8B scale.
SPHINX reveals that understanding a policy's failures leads to more effective adversarial scenario generation, enhancing the robustness of autonomous driving systems.
Forget relying on decoder vectors or LLMs for SAE feature analysis – Semantic Optimal Transport offers a provably robust and empirically superior way to match features and compress circuits by modeling features as distributions over hidden states.
