Poser: Unmasking Alignment Faking LLMs by Manipulating Their Internals

Joshua Clymer, Caden Juang, Severin Field
5/8/2024
cs.CLcs.AI

No official code found. Best implementation: sevdeawesome/POSER (7 stars, last commit 2024-08-27).

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.

View on arXivView PDF

Code Implementations(1)

sevdeawesome/POSERCommunity100%

Poser: Unmasking Alignment Faking LLMs by Manipulating Their Internals

74Shell, PythonMay 20, 20242 years ago

Similar Papers

Empirical Evidence for Alignment Faking in a Small LLM and Prompt-Based Mitigation Techniques

88% match2025

Alignment faking in large language models

85% match2 repos952024

When Thinking LLMs Lie: Unveiling the Strategic Deception in Representations of Reasoning Models

84% match2025

Why Do Some Language Models Fake Alignment While Others Don't?

83% match2025

Aligners: Decoupling LLMs and Alignment

83% match1 repo2024

Can LLMs Lie? Investigation beyond Hallucination

82% match2025

Cite this paper

@article{clymer2024poser,
  title  = {Poser: Unmasking Alignment Faking LLMs by Manipulating Their Internals},
  author = {Joshua Clymer and Caden Juang and Severin Field},
  year   = {2024},
  eprint = {2405.05466},
  archivePrefix = {arXiv},
  url    = {http://arxiv.org/abs/2405.05466v2}
}

Add a badge to your README

Shows whether official code is linked for this paper.

Code of Paper badge
[![Code of Paper](https://codeofpaper.com/badge/2405.05466.svg)](https://codeofpaper.com/paper/2405.05466)

Discussion