Pooyan Rahmanzadehgervi

CS PhD candidate @AuburnU

[Pooyan Rahmanzadehgervi @Pooyanrg]

Interesting to think about TAB through the lens of attention sinks as mechanisms rather than patterns. 8/8

[Pooyan Rahmanzadehgervi @Pooyanrg]

I really enjoyed reading this amazing paper from @thomas_fel_ et al. 🙌 Great read!
https://t.co/fRDrBmuuAo
1/8

[Pooyan Rahmanzadehgervi @Pooyanrg]

When differences are present, we force the bottleneck to localize the changed region, making it a task-relevant and interpretable channel for communicating (broadcasting) object-level differences. 7/8

[Tin (Kevin) Nguyen @tin_ng_qn]

Reading LLM answers but don’t know where they come from?
Struggling to find the right button to click?
Or distracted by cluttered content on the page?

We built PageGuide (https://t.co/WdLKkiZym7) to fix that

[Logan Bolton @septisum]

Hood popped. Photo taken. “Hey ChatGPT, how do I check my car’s oil level?” And it returns a giant block of text…

Instead, a human would point to the oil cap and draw on the photo to answer! We explore how to unlock VLMs to do that, i.e., annotating on the image to guide users through answers visually: https://t.co/nOgC7XuW4z

1/n 🧵

[Pooyan Rahmanzadehgervi @Pooyanrg]

@septisum @taesiri @anh_ng8 Gemini 3 Pro preview❌
(2/2)

Model thoughts:

[Pooyan Rahmanzadehgervi @Pooyanrg]

A year and half since or team (@septisum, @taesiri, @anh_ng8 ) introduced BlindTest (https://t.co/ZFV4662YDv).

Some million dollars and many thinking tokens spent and still:

[Pooyan Rahmanzadehgervi @Pooyanrg]

@septisum @taesiri @anh_ng8 Gemini 3 Pro preview ❌
(1/2)

[Pooyan Rahmanzadehgervi @Pooyanrg]

At this moment, ICLR rebuttal vibe be like 🤣

[Pooyan Rahmanzadehgervi @Pooyanrg]

- TAB: Transformer Attention Bottlenecks enable User Intervention and Debugging in Vision-Language Models (https://t.co/ysKHzEEcxL)
- Vision language models fail to translate detailed visual features into words (https://t.co/wkxddpOJlS)

[Pooyan Rahmanzadehgervi @Pooyanrg]

I will present 2 posters today @ eXCV Workshop at @ICCVConference.

Please stop by between 11:15 - 14:00 (HST) in Exhibit Hall 2 at posters 14 and 15 to chat if you are in Honolulu.

[Hokin Deng @DengHokin]

#embodied All forms of biological intelligence are grounded movements🏃‍♂️ muscles & motor neurons 🧠 emerge before visual cortex & rods & cones in eyes 👁️

Building monocular better-than-mocap-studio #video2motion is our critical step towards human embodied intelligence.

[Pooyan Rahmanzadehgervi @Pooyanrg]

I have started my PhD with TAB, and it has finally paid off!

Code: https://t.co/WE5eTCcuOo

[Pooyan Rahmanzadehgervi @Pooyanrg]

Happy to share that "TAB: Transformer Attention Bottlenecks enable User Intervention and Debugging in Vision-Language Models" has been accepted to #ICCV 2025. 🔥

Shout out to amazing collaborators:
Hung H. Nguyen, @savvyRL , Long Mai, and @anh_ng8. 🤩

https://t.co/x2L5CNXE5q

[Anh Totti Nguyen @anh_ng8]

Pooyan @Pooyanrg presenting our Transformer Attention Bottleneck paper at @CVPR

💡 We **simplify** MHSA (e.g. 12 heads -> 1 head) to create an attention **bottleneck** where users can debug Vision Language Models by editing the bottleneck and observe expected VLM text outputs.