Simple LLM Baselines are Competitive for Model Diffing

Elias Kempf, Simon Schrodi, Bartosz Cywiński, Thomas Brox, Neel Nanda +1 more
2/10/2026
cs.LG

No official code found. Best implementation: eliaskempf/model-diffing (10 stars, MIT, last commit 2026-02-13).

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.

View on arXivView PDF

Code Implementations(1)

eliaskempf/model-diffingCommunity62%
MIT

Official code for the paper: "Simple LLM Baselines are Competitive for Model Diffing"

100Shell, PythonFeb 10, 20268 months agoMIT
black-boxhypothesis-generationinterpretabilitylanguage-modelsmodel-diffing

Similar Papers

Enhancing LLM Evaluations: The Garbling Trick

85% match2024

No LLM is Free From Bias: A Comprehensive Study of Bias Evaluation in Large Language Models

82% match2025

Reviving Your MNEME: Predicting The Side Effects of LLM Unlearning and Fine-Tuning via Sparse Model Diffing

82% match2025

Rating Roulette: Self-Inconsistency in LLM-As-A-Judge Frameworks

82% match2025

Prompt Variability Effects On LLM Code Generation

82% match2025

Estimating LLM Consistency: A User Baseline vs Surrogate Metrics

81% match2025

Cite this paper

@article{kempf2026simple,
  title  = {Simple LLM Baselines are Competitive for Model Diffing},
  author = {Elias Kempf and Simon Schrodi and Bartosz Cywiński and Thomas Brox and Neel Nanda and Arthur Conmy},
  year   = {2026},
  eprint = {2602.10371},
  archivePrefix = {arXiv},
  url    = {http://arxiv.org/abs/2602.10371v1}
}

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/2602.10371.svg)](https://codeofpaper.com/paper/2602.10371)

Discussion