VFUSE applies sparse autoencoders to diffusion-transformer activations in RoseTTAFold3 and RFDiffusion3 to find monosemantic features that detect hazardous protein designs with AUROC up to 0.84.
Aaron Maiwald, Piotr Jedryszek, Florent Draye, Garrett M
3 Pith papers cite this work. Polarity classification is still indexing.
3
Pith papers citing it
years
2026 3verdicts
UNVERDICTED 3representative citing papers
ESM2 predicts N-terminal methionine via retrieval of a positional prior from the BOS token through distributed attention circuits rather than direct recognition, revealed by a norm-direction decomposition of rotary attention scores.
ESMC-SAE features enable 78.9% top-1 EC number prediction on 4,868 microbial enzymes, outperforming 3-mer baselines by 37.6% and recovering novel EC1 classes at 47.7% in leave-one-out tests.
citing papers explorer
-
VFUSE: Virulent Feature Understanding with Sparse autoEncoders
VFUSE applies sparse autoencoders to diffusion-transformer activations in RoseTTAFold3 and RFDiffusion3 to find monosemantic features that detect hazardous protein designs with AUROC up to 0.84.
-
Retrieval and competition: how a protein foundation model starts a protein
ESM2 predicts N-terminal methionine via retrieval of a positional prior from the BOS token through distributed attention circuits rather than direct recognition, revealed by a norm-direction decomposition of rotary attention scores.
-
Interpretable enzyme function prediction via sparse autoencoder features of ESMC across the microbial protein universe
ESMC-SAE features enable 78.9% top-1 EC number prediction on 4,868 microbial enzymes, outperforming 3-mer baselines by 37.6% and recovering novel EC1 classes at 47.7% in leave-one-out tests.