Matt Gorbett

Cross-Model Disagreement as a Label-Free Correctness Signal

Matt Gorbett, Suman Jana

Secure Linear Alignment of Large Language Models

Matt Gorbett, Suman Jana

This paper shows that different language models learn surprisingly similar internal representations, even when trained differently. The researchers exploit this similarity to create a privacy-preserving system where one model can use another model's knowledge through a simple mathematical transformation (affine alignment), without sharing raw data or model weights—using encryption to keep queries secure.