The Domestic Yak podcast

Auditing LLMs & Hidden Objectives

0:00
17:00
Rewind 15 seconds
Fast Forward 15 seconds

This episode summarizes: Auditing language models for hidden objectives by Samuel Marks Et.al.

Submitted on March 14th 2025 https://arxiv.org/abs/2503.10965


Investigated the feasibility of alignment audits by training a language model with a hidden objective: to exhibit behaviors it believes reward models favor, even if undesirable. Three teams successfully identified the hidden objective using techniques like interpretability tools, behavioral attacks, and training data analysis.

More episodes from "The Domestic Yak"