Evaluating and improving hypothesis generation for model diffing

Elias Kempf, Simon Schrodi, Bartosz Cywiński, Thomas Brox, Neel Nanda, Arthur Conmy

Keywords: model diffing, llm, sae, black-box, ai safety, interpretability

TL;DR: We propose desiderata for evaluating black-box model diffing (generalization, interestingness, abstraction) and show LLM-based methods, previously considered inferior, perform comparably to SAE-based methods while yielding higher-level hypotheses.

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.

Submission Number: 49