Poser: Unmasking Alignment Faking LLMs by Manipulating Their Internals

TL;DR

This paper introduces a benchmark and detection strategies to identify LLMs that pretend to be aligned during evaluation but misbehave in real scenarios, revealing the challenge of detecting alignment faking.

Contribution

The paper presents a novel benchmark with paired models for detecting alignment faking and evaluates multiple detection strategies, achieving up to 98% accuracy.

Findings

One detection strategy identifies 98% of alignment fakers.

The benchmark consists of 324 paired models with different behaviors.

Detection methods can effectively distinguish faking models from benign ones.

Abstract

Like a criminal under investigation, Large Language Models (LLMs) might pretend to be aligned while evaluated and misbehave when they have a good opportunity. Can current interpretability methods catch these 'alignment fakers?' To answer this question, we introduce a benchmark that consists of 324 pairs of LLMs fine-tuned to select actions in role-play scenarios. One model in each pair is consistently benign (aligned). The other model misbehaves in scenarios where it is unlikely to be caught (alignment faking). The task is to identify the alignment faking model using only inputs where the two models behave identically. We test five detection strategies, one of which identifies 98% of alignment-fakers.