Simple LLM Baselines are Competitive for Model Diffing

TL;DR

This paper compares different model diffing methods for large language models, proposing evaluation metrics and showing that an improved LLM-based baseline is competitive with SAE-based methods in surfacing behavioral differences.

Contribution

It introduces evaluation criteria for model diffing methods and systematically compares LLM-based and SAE-based approaches.

Findings

LLM-based baseline performs comparably to SAE-based method.

The improved LLM-based method surfaces more abstract behavioral differences.

Evaluation metrics for model diffing are proposed and validated.

Abstract

Standard LLM evaluations only test capabilities or dispositions that evaluators designed them for, missing unexpected differences such as behavioral shifts between model revisions or emergent misaligned tendencies. Model diffing addresses this limitation by automatically surfacing systematic behavioral differences. Recent approaches include LLM-based methods that generate natural language descriptions and sparse autoencoder (SAE)-based methods that identify interpretable features. However, no systematic comparison of these approaches exists nor are there established evaluation criteria. We address this gap by proposing evaluation metrics for key desiderata (generalization, interestingness, and abstraction level) and use these to compare existing methods. Our results show that an improved LLM-based baseline performs comparably to the SAE-based method while typically surfacing more abstract behavioral differences.