Twitxy Saved

Babak Ehteshami Bejnordi

@BabakEht

Principal Research Scientist @Nvidia | Nemotron
Previously: @Qualcomm AI Research
Zurich, Switzerland babakint.github.io Joined October 2012
283 Following 356 Followers 114 Posts

Latest posts 5

Babak Ehteshami Bejnordi @BabakEht
I am attending #ICLR in Rio de Janeiro this week 🇧🇷

If you’re interested in efficient LLM inference, latent reasoning, or MoEs, I’d be happy to meet and chat at our posters:

Come find @a_kzna, @maciejpioro and me here! 👇
On my way to ICLR to present our work on Latent Reasoning.

Ping me here if you want to grab a coffee or drop by our poster!

Sat, Apr 25, 2026, 3:15 – 5:45 PM
📍 Pavilion 3 P3-#905
View quoted post on X
Babak Ehteshami Bejnordi @BabakEht
Introducing KaVa (arxiv.org/abs/2510.02312): a latent reasoning framework that teaches a model to think efficiently. It’s the first to use a teacher’s compressed KV-cache as a direct supervisory signal for a latent-reasoning student. Fantastic joint work w. @maciejpioro & @a_kzna.
arxiv.org KaVa: Latent Reasoning via Compressed KV-Cache Distillation Large Language Models (LLMs) excel at multi-step reasoning problems with explicit chain-of-thought (CoT), but verbose traces incur significant computational costs and memory overhead, and often...
Maciej Pióro 𐙼 Verified account @maciejpioro Oct 3, 2025
(1/n) Introducing KaVa (arxiv.org/abs/2510.02312) – the first latent reasoning framework leveraging compressed KV-Cache to guide the latent generation. We beat previous approaches, especially on a realistic, Natural Language GSM8K dataset.
View quoted post on X
Babak Ehteshami Bejnordi @BabakEht
We propose a dynamic tokenizer for ViTs, where the scale at which an image is processed varies based on the complexity of the image area. This means less computing for simple areas and more for complex, cluttered areas. Thanks to @royaleerieme, @JakobHavtorn, @TiRune
AK Verified account @_akhaliq Jul 6, 2023
MSViT: Dynamic Mixed-Scale Tokenization for Vision Transformers

paper page: huggingface.co/papers/2307.02…

The input tokens to Vision Transformers carry little semantic meaning as they are defined as regular equal-sized patches of the input image, regardless of its content. However, processing uniform background areas of an image should not necessitate as much compute as dense, cluttered areas. To address this issue, we propose a dynamic mixed-scale tokenization scheme for ViT, MSViT. Our method introduces a conditional gating mechanism that selects the optimal token scale for every image region, such that the number of tokens is dynamically determined per input. The proposed gating module is lightweight, agnostic to the choice of transformer backbone, and trained within a few epochs (e.g., 20 epochs on ImageNet) with little training overhead. In addition, to enhance the conditional behavior of the gate during training, we introduce a novel generalization of the batch-shaping loss. We show that our gating module is able to learn meaningful semantics despite operating locally at the coarse patch-level. We validate MSViT on the tasks of classification and segmentation where it leads to improved accuracy-complexity trade-off.
View quoted post on X
0 replies 8 reposts 45 likes 17K views View on X