Phongsakon Mark Konrad

New method audits fine-tuned language model deployment mismatches

Researchers have developed a new method called "Routing Subspaces" to identify and address the mismatch between how fine-tuned language models perform during safety evaluations and how they behave in real-world deployme…