https://xitter.cf/t.co/srLqFNHBVr Previously: Mechanistic Interpretability for Vision (MIV) Workshop
Joined January 2025
- Tweets84
- Following17
- Followers302
- Likes14
We're back for @CVPR 2026!
Meet ✨HOW 2026: How Do Vision Models Work?✨ workshop
Building on the success of last year’s MIV workshop, we are expanding our lens beyond mechinterp to build a unified science of how vision models learn, represent, and reason.
More details 👇
How Do Vision Models Work? @ CVPR2026 (Prev: MIV) retweeted
Large-scale automated neuron labeling for... neurons in the brain! Excited to see ideas from AI interpretability helping us understand biological neural systems. One result I love: 96% of V4 neurons are language-describable! Fortunate to be part of this project led by @vedanglad
How Do Vision Models Work? @ CVPR2026 (Prev: MIV) retweeted
We just revamped the Gaze Heads demo with two powerful new features:
🔍 Scroll-to-zoom
✨ Continuous hover
Direct your VLM’s gaze attention to any region, anytime - and get a description in the exact order you point !!
Try it: gaze.baulab.info/#demo 👇
cc: @davidbau
Vision-language AI models have a gaze. And you can steer it! 👀
Redirect just 9% of a model’s attention heads to any region in an image, and the VLM will start describing that region mid-generation. We call them Gaze Heads!
Try the demo: gaze.baulab.info/#demo 🧵👇
How Do Vision Models Work? @ CVPR2026 (Prev: MIV) retweeted
Vision-language AI models have a gaze. And you can steer it! 👀
Redirect just 9% of a model’s attention heads to any region in an image, and the VLM will start describing that region mid-generation. We call them Gaze Heads!
Try the demo: gaze.baulab.info/#demo 🧵👇
How Do Vision Models Work? @ CVPR2026 (Prev: MIV) retweeted
the workshop admin is so hype about NDIF, love it!
yes dw about GPUs
