Simple LLM Baselines are Competitive for Model Diffing

TL;DR

This work investigates how to systematically surface unexpected behavioral differences between API-accessible LLMs through model diffing, addressing gaps in conventional evaluations that focus on specific capabilities or dispositions. It compares an improved LLM-based baseline with a sparse autoencoder (SAE)–based method across data from WildChat prompts, using a judge-based verification framework and metrics for generalization, interestingness, and abstraction. Generalization is formalized via $f(h)$ and $acc(h)$, with the derived $vfd(h)$ capturing the trade-off between how often a behavior appears and how reliably it discriminates between models. Across three experiments, the LLM-based method matches the SAE-based method on accuracy and frequency but yields more abstract hypotheses and higher acceptance rates, suggesting that simple LLM baselines remain strong for API-only model diffing and can reveal high-level behavioral shifts relevant for safety and deployment. The study highlights practical implications, acknowledges limitations of automatic evaluation, and points to promising future directions, including hybrid approaches and targeted prompt strategies to uncover domain-specific or reasoning-level differences.

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.

Figures (17)

• Figure 1: Model diffing and evaluation pipeline.\ref{['fig:main:diffing']}: To understand the differences between two API-only models, we collect completions from both models to a joint set of prompts, and generate hypotheses about behavioral differences between the two models using either LLM- or SAE-based API-only model diffing methods. \ref{['fig:main:results']}: We compare the quality of different model diffing results using a set of five evaluation metrics: accuracy, frequency, interestingness, abstraction level, and acceptance rate (see \ref{['sec:evaluation']} for details).
• Figure 2: Evaluation results. Across our experiments, the LLM- and SAE-based method perform very similar in terms of accuracy, frequency, and interestingness of generated hypotheses. The LLM-based method provides hypotheses at a consistently higher abstraction level while usually also providing a higher acceptance rate.
• Figure 3: Response lengths.
• Figure 4: System prompt for LLM judge.
• Figure 5: User prompt for LLM judge.