To summarize our experimental results, we propose a new metric that quantifies how well a method isolates model differences. We found RDX consistently outperforms baseline approaches on this metric and several other established metrics.

Beyond controlled experiments, RDX uncovers previously unknown differences. For example, we found DINOv2 has extra structure for distinguishing monkey species, helping explain its improved fine-grained performance.

In a case study on models with small performance differences, baselines (like SAE and NMF) mostly captured shared concepts. RDX, in contrast, localized groups of images that were incorrectly clustered in the weaker model, revealing subtle but important differences.

In the game, you’ll see that prior methods often highlight the wrong parts of each model’s representations by explaining shared structure. RDX, by contrast, consistently focuses on the differences between models, making it much easier to interpret what changed during training. 🔍

Excited to share our paper Representational Difference Explanations (RDX) was accepted to #NeurIPS2025! 🎉RDX is a new method for model diffing designed to isolate 🔍 representational differences. 1/7

Even on a simple MNIST model, it is essentially impossible to anticipate that a weighted sum over these explanations results in this normal-looking five. Linear combinations of explanation grids are tricky to understand!

We compare RDX to several popular dictionary-learning (DL) methods (like SAEs and NMF) and find that the DL methods struggle. In the spotted wing (SW) comparison experiment, we find that NMF shows model similarities rather than differences.

After demonstrating that RDX works when there are known differences, we compare models with unknown differences. For example, when comparing DINO and DINOv2, we find that DINOv2 has learned a color based categorization of gibbons that is not present in DINO.

We apply RDX on trained models with known differences and show that it isolates the core differences. For example, we compare model representations with and w/out a “spotted wing” (SW) concept and find that RDX shows that only one model groups birds according to this feature.

We found these unique and important concepts to be fairly complex, requiring deep analysis. We use ChatGPT-4o to analyze the concept collages and find that it gives detailed and clear explanations about the differences between models. More examples here -- nkondapa.github.io/rsvc-page/

We then look at “in-the-wild” models. We compare ResNets and ViTs trained on ImageNet. We measure concept importance and concept similarity. Do models learn unique and important concepts? Yes, sometimes they do!

We first show this approach can recover known differences. We train Model 1 to use a pink square to make classification decisions and Model 2 to ignore it. Our method, RSVC, isolates this difference.

Have you ever wondered what makes two models different?
We all know the ViT-Large performs better than the Resnet-50, but what visual concepts drive this difference? Our new ICLR 2025 paper addresses this question! nkondapa.github.io/rsvc-page/