Just a quick update: the full text of our CVPR Findings 2026 paper, "Large Multimodal Models as General In-Context Classifiers", is now officially available as a preprint on arXiv.

If you want to dive straight into the technical details behind our benchmarking, the open-world challenges of large multimodal models (LMMs), and the exact mechanics of our training-free CIRCLE method, the preprint is already 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!

I am thrilled to announce that our latest paper, investigating the true classification potential of large multimodal models (LMMs), has been accepted to CVPR Findings 2026!

For a long time, the computer vision community has relied on a generally accepted rule of thumb: use CLIP-like contrastive vision-language models (VLMs) for zero-shot classification, and save LMMs for more complex, generative tasks.

In our new paper, we challenge that status quo.

We wanted to know what happens when you leverage a frequently overlooked superpower of LMMs: in-context learning. Here is a brief look at what we discovered:

Through extensive experiments, we demonstrated that CIRCLE establishes a highly robust baseline for open-world classification. It consistently outperforms VLM equivalents, proving that large multimodal models have the potential to act as powerful, unified classifiers and flexible alternatives to highly specialized models.

I can't wait to share more details with the community at CVPR 2026. Stay tuned for the full paper release and the accompanying code!

Starting today, ComCa is openly available on arXiv, waiting to be read by all researchers in the community of open-vocabulary detection, working in training-free, retrieval-based settings, and specifically attribute detection.

The code has been developed with the latest libraries employed in deep learning, such as:

Code is available here.

This research leverages pre-trained vision-language models (VLMs) for the task of attribute detection, which involves identifying all attributes of an object, such as color, material and texture.

ComCa, our training-free method, surpasses all other training-free models and is even competitive with training-based approaches.

By introducing a caching strategy that stores examples of objects with specific attributes, ComCa can refine the VLM's predictions at test time requiring no specific training nor manually annotated data. Indeed, the cache is constructed automatically by leveraging web-scale databases, such as CC12M, and large language models (LLMs) such as GPT by OpenAI.

Code will be released shortly, and the camera-ready version of the paper will be uploaded on arXiv as soon as it is ready.

If you are interested in COPS, take a look at it!