Replicated "Physics of LMs: Part 3.1" (Allen-Zhu & Li) arxiv.org/pdf/2309.14316, no official code existed.
Core finding: memorization ≠ knowledge extraction. Near-perfect pretraining loss, 0% QA accuracy. Data augmentation is key to extract knowledge

Code : github.com/ykrmm/Physic...
PRs welcome.

https://aclanthology.org/2025.acl-long.919.pdf

⚡ I am starting my postdoctoral research in the ALMANACH team at Inria 🇫🇷 and Inria Chile on the topic of Knowledge-Graph-Augmented LLMs for Cultural References!

Current LLMs often struggle with cultural references of underrepresented groups, leading to strong biases and hallucinations.

GitHub - ykrmm/ViLU: ICCV'25 ViLU: Learning Vision-Language Uncertainties for Failure Prediction ICCV'25 ViLU: Learning Vision-Language Uncertainties for Failure Prediction - ykrmm/ViLU

Link to the code ⌨️ :

github.com/ykrmm/ViLU

ViLU: Learning Vision-Language Uncertainties for Failure Prediction Reliable Uncertainty Quantification (UQ) and failure prediction remain open challenges for Vision-Language Models (VLMs). We introduce ViLU, a new Vision-Language Uncertainty quantification framework ...

Link to the paper 📝 :
arxiv.org/abs/2507.07620

🎉 Our work on uncertainty quantification in Vision-Language Models has been accepted at ICCV! 🌴

We propose a simple yet effective post-hoc model that learns to capture ambiguity from both images and captions, achieving high accuracy in detecting VLM failures.

🚀Thrilled to introduce JAFAR—a lightweight, flexible, plug-and-play module that upsamples features from any Foundation Vision Encoder to any desired output resolution (1/n)

Paper : arxiv.org/abs/2506.11136
Project Page: jafar-upsampler.github.io
Github: github.com/PaulCouairon...

1/n 🚀New paper out - accepted at #ICCV2025!

Introducing DIP: unsupervised post-training that enhances dense features in pretrained ViTs for dense in-context scene understanding

Below: Low-shot in-context semantic segmentation examples. DIP features outperform DINOv2!

Well arrived in Vancouver 🇨🇦 ! Can't wait to start the NeurIPS conference!

Léo presenting his article

A group of musicians playing for the jam session

San Francisco by night

The Golden Gate Bridge

Two weeks ago, I had the chance to attend the #ISMIR2024 conference in San Francisco! I presented our work on hierarchical representations of music for classification models: arxiv.org/abs/2407.17536

It was an amazing experience, with great conversations and great people! (and talented musicians!)

Supra-Laplacian Encoding for Transformer on Dynamic Graphs Fully connected Graph Transformers (GT) have rapidly become prominent in the static graph community as an alternative to Message-Passing models, which suffer from a lack of expressivity, oversquashing...

Are you interested in dynamic graphs and transformers ?

Check out our latest paper accepted at NeurIPS 🥳 !

We desgined a new spatio-temporal encoding based on the spectral properties of a supra-laplacian matrix associated to a dynamic graph.

Code is coming soon !

arxiv.org/abs/2409.17986

Hello everyone!

My name is Yannis Karmim, I'm a phd student at Conservatoire National des Arts et Métiers in Paris 🇫🇷!
I'm working on representation learning on dynamic graphs.

I'm also interested in VLM uncertainty in parallel to my PhD.

To find out more about my work --> ykrmm.github.io

Congrats Georges !

GitHub - gle-bellier/flow-matching: Annotated Flow Matching paper Annotated Flow Matching paper. Contribute to gle-bellier/flow-matching development by creating an account on GitHub.

I created 3 introductory notebooks on Flow Matching models to help get started with this exciting topic! ✨

1. Annotated Flow Matching paper: github.com/gle-bellier/...
2. Discrete Flow Matching: github.com/gle-bellier/...
3. Minimal FM in Jax: github.com/gle-bellier/...