Project page and code for CIRCLE are available

Following up on the recent acceptance of our paper to CVPR Findings 2026, I am excited to share that the official project page and the full codebase for our research are now publicly available!

If you read my previous post, you know we have been investigating how large multimodal models (LMMs) can challenge contrastive vision-language models (VLMs) like CLIP in classification tasks. By leveraging in-context learning, we found that LMMs can serve as powerful, unified classifiers - especially in open-world scenarios.

To make our findings as accessible and reproducible as possible, we have open-sourced everything.

We have set up a dedicated project page. Take a look at it - it has some cool animations that help you understand the method!

We also released the code on GitHub, at this page. The code contains the implementation of CIRCLE, ready to use out-of-the-box, together with evaluation scripts and documentation.

If you find our work useful, please consider starring the repository!