BioinformaticsbioRxivPreprint
The Confessions of a Protein Language Model: Steering, Activations, and What Predicted Confidence Cannot Certify
Panini Shah, Raj Panchali, Pearl Mody…·Ruhina Karani
The paper audits protein generate-and-screen pipelines by testing steering, confidence, and activation readouts, finding they do not reliably predict foldability.
DOI 10.64898/2026.11.10.765136