Evangelos Kazakos
@ekazakos.bsky.social
Postdoctoral researcher @ CIIRC, CTU, Prague working in vision & language. Also robotics noob. PhD from University of Bristol. Ex. Samsung Research (SAIC-C). I love coffee and plants. And socks.
Pinned
Evangelos Kazakos
@ekazakos.bsky.social
· Jun 26
🚨🚨📢📢 PAPER ALERT 📢📢🚨🚨
Excited to announce our new work: "Large-scale Pre-training for Grounded Video Caption Generation" with Cordelia Schmid & @josef-sivic.bsky.social.
Paper: arxiv.org/abs/2503.10781
Project: ekazakos.github.io/grounded_vid...
Code (coming soon): github.com/ekazakos/grove 1/7
Excited to announce our new work: "Large-scale Pre-training for Grounded Video Caption Generation" with Cordelia Schmid & @josef-sivic.bsky.social.
Paper: arxiv.org/abs/2503.10781
Project: ekazakos.github.io/grounded_vid...
Code (coming soon): github.com/ekazakos/grove 1/7
Large-scale Pre-training for Grounded Video Caption Generation
We propose a novel approach for captioning and object grounding in video, where the objects in the caption are grounded in the video via temporally dense bounding boxes. We introduce the following con...
arxiv.org
Our work, GROVE, has been accepted to ICCV 2025! 🎉 This is collab. w. Cordelia Schmid & @josef-sivic.bsky.social.
We will release code, models and datasets within next 2 weeks.
We are also working on a search demo for the proposed datasets with user prompts!
I hope to see you all in Honolulu!
We will release code, models and datasets within next 2 weeks.
We are also working on a search demo for the proposed datasets with user prompts!
I hope to see you all in Honolulu!
Reposted by Evangelos Kazakos
Found myself repeating this to several students at NeurIPS:
When you’re choosing an internship or a job, what you work on and who you work with matter way more than the logo. Don’t optimize for brands. Become the brand!
When you’re choosing an internship or a job, what you work on and who you work with matter way more than the logo. Don’t optimize for brands. Become the brand!
December 9, 2025 at 2:17 AM
Found myself repeating this to several students at NeurIPS:
When you’re choosing an internship or a job, what you work on and who you work with matter way more than the logo. Don’t optimize for brands. Become the brand!
When you’re choosing an internship or a job, what you work on and who you work with matter way more than the logo. Don’t optimize for brands. Become the brand!
Reposted by Evangelos Kazakos
Historian Thomas Hughes argued that technologies are malleable when young, then harden. Right now we're still shaping AI, or at least it is being shaped by our institutions, norms & use cases
Eventually these systems build a momentum of their own. That is why choices now matter, things are fluid.
Eventually these systems build a momentum of their own. That is why choices now matter, things are fluid.
December 9, 2025 at 12:20 AM
Historian Thomas Hughes argued that technologies are malleable when young, then harden. Right now we're still shaping AI, or at least it is being shaped by our institutions, norms & use cases
Eventually these systems build a momentum of their own. That is why choices now matter, things are fluid.
Eventually these systems build a momentum of their own. That is why choices now matter, things are fluid.
Reposted by Evangelos Kazakos
TITANS & MIRAS: real continual learning
MIRAS = a unifying theory of transformers (attention) and state space models (SSM, e.g. Mamba, RNNs)
TITANS = an optimal MIRAS implementation that’s “halfway between” SSM & transformer with a CL memory module
let’s dive in!
research.google/blog/titans-...
MIRAS = a unifying theory of transformers (attention) and state space models (SSM, e.g. Mamba, RNNs)
TITANS = an optimal MIRAS implementation that’s “halfway between” SSM & transformer with a CL memory module
let’s dive in!
research.google/blog/titans-...
Titans + MIRAS: Helping AI have long-term memory
research.google
December 6, 2025 at 4:04 PM
TITANS & MIRAS: real continual learning
MIRAS = a unifying theory of transformers (attention) and state space models (SSM, e.g. Mamba, RNNs)
TITANS = an optimal MIRAS implementation that’s “halfway between” SSM & transformer with a CL memory module
let’s dive in!
research.google/blog/titans-...
MIRAS = a unifying theory of transformers (attention) and state space models (SSM, e.g. Mamba, RNNs)
TITANS = an optimal MIRAS implementation that’s “halfway between” SSM & transformer with a CL memory module
let’s dive in!
research.google/blog/titans-...
Reposted by Evangelos Kazakos
#NeurIPS2026 will be held in Sydney, Australia!
#ICML2017 was also in Sydney and was an absolute blast
#ICML2017 was also in Sydney and was an absolute blast
December 6, 2025 at 3:41 AM
#NeurIPS2026 will be held in Sydney, Australia!
#ICML2017 was also in Sydney and was an absolute blast
#ICML2017 was also in Sydney and was an absolute blast
Reposted by Evangelos Kazakos
🏹 Job alert: Explainable AI Methods for Self-Driving Labs at Karlsruhe Institute of Technology
📍 Karlsruhe 🇩🇪
📅 Apply by Dec 31st
🔗 https://bit.ly/49NJsWJ
📍 Karlsruhe 🇩🇪
📅 Apply by Dec 31st
🔗 https://bit.ly/49NJsWJ
KIT - PSE - For applicants - Jobs - Joboffer-Details
bit.ly
December 5, 2025 at 3:05 PM
🏹 Job alert: Explainable AI Methods for Self-Driving Labs at Karlsruhe Institute of Technology
📍 Karlsruhe 🇩🇪
📅 Apply by Dec 31st
🔗 https://bit.ly/49NJsWJ
📍 Karlsruhe 🇩🇪
📅 Apply by Dec 31st
🔗 https://bit.ly/49NJsWJ
Reposted by Evangelos Kazakos
🏹 Job alert: Postdocs on diversity and inclusion in machine learning competitions at IT University of Copenhagen
📍 Copenhagen 🇩🇰
📅 Apply by Jan 31st
🔗 https://candidate.hr-manager.net/ApplicationInit.aspx?cid=119&ProjectId=181837&DepartmentId=3439&MediaId=5
📍 Copenhagen 🇩🇰
📅 Apply by Jan 31st
🔗 https://candidate.hr-manager.net/ApplicationInit.aspx?cid=119&ProjectId=181837&DepartmentId=3439&MediaId=5
Postdocs on diversity and inclusion in machine learning competitions
The PURRlab (Pattern Recognition Revisited lab) at the IT University of Copenhagen invites motivated individuals to apply for postdoctoral positions starting in
candidate.hr-manager.net
December 4, 2025 at 5:06 PM
🏹 Job alert: Postdocs on diversity and inclusion in machine learning competitions at IT University of Copenhagen
📍 Copenhagen 🇩🇰
📅 Apply by Jan 31st
🔗 https://candidate.hr-manager.net/ApplicationInit.aspx?cid=119&ProjectId=181837&DepartmentId=3439&MediaId=5
📍 Copenhagen 🇩🇰
📅 Apply by Jan 31st
🔗 https://candidate.hr-manager.net/ApplicationInit.aspx?cid=119&ProjectId=181837&DepartmentId=3439&MediaId=5
Reposted by Evangelos Kazakos
𝗜𝗻𝘁𝗲𝗿𝗻𝗩𝗶𝗱𝗲𝗼-𝗡𝗲𝘅𝘁: 𝗧𝗼𝘄𝗮𝗿𝗱𝘀 𝗚𝗲𝗻𝗲𝗿𝗮𝗹 𝗩𝗶𝗱𝗲𝗼 𝗙𝗼𝘂𝗻𝗱𝗮𝘁𝗶𝗼𝗻 𝗠𝗼𝗱𝗲𝗹𝘀 𝘄𝗶𝘁𝗵𝗼𝘂𝘁 𝗩𝗶𝗱𝗲𝗼-𝗧𝗲𝘅𝘁 𝗦𝘂𝗽𝗲𝗿𝘃𝗶𝘀𝗶𝗼𝗻
Chenting Wang, Yuhan Zhu, Yicheng Xu ... Limin Wang
arxiv.org/abs/2512.01342
Trending on www.scholar-inbox.com
Chenting Wang, Yuhan Zhu, Yicheng Xu ... Limin Wang
arxiv.org/abs/2512.01342
Trending on www.scholar-inbox.com
December 4, 2025 at 7:00 AM
𝗜𝗻𝘁𝗲𝗿𝗻𝗩𝗶𝗱𝗲𝗼-𝗡𝗲𝘅𝘁: 𝗧𝗼𝘄𝗮𝗿𝗱𝘀 𝗚𝗲𝗻𝗲𝗿𝗮𝗹 𝗩𝗶𝗱𝗲𝗼 𝗙𝗼𝘂𝗻𝗱𝗮𝘁𝗶𝗼𝗻 𝗠𝗼𝗱𝗲𝗹𝘀 𝘄𝗶𝘁𝗵𝗼𝘂𝘁 𝗩𝗶𝗱𝗲𝗼-𝗧𝗲𝘅𝘁 𝗦𝘂𝗽𝗲𝗿𝘃𝗶𝘀𝗶𝗼𝗻
Chenting Wang, Yuhan Zhu, Yicheng Xu ... Limin Wang
arxiv.org/abs/2512.01342
Trending on www.scholar-inbox.com
Chenting Wang, Yuhan Zhu, Yicheng Xu ... Limin Wang
arxiv.org/abs/2512.01342
Trending on www.scholar-inbox.com
Reposted by Evangelos Kazakos
We have publicly shared our "PowerCLIP," a method to align powersets of image sub-region with textual structures for precise image-text recognition.
Outperforms several SotA in zero-shot classification, retrieval, robustness, and compositional tasks!
arxiv.org/abs/2511.23170
Outperforms several SotA in zero-shot classification, retrieval, robustness, and compositional tasks!
arxiv.org/abs/2511.23170
December 4, 2025 at 4:20 AM
We have publicly shared our "PowerCLIP," a method to align powersets of image sub-region with textual structures for precise image-text recognition.
Outperforms several SotA in zero-shot classification, retrieval, robustness, and compositional tasks!
arxiv.org/abs/2511.23170
Outperforms several SotA in zero-shot classification, retrieval, robustness, and compositional tasks!
arxiv.org/abs/2511.23170
Reposted by Evangelos Kazakos
Super-exciting talk by Ani Kembhavi from Wayve AI @bristoluni.bsky.social @compscibristol.bsky.social #MaVi Seminar today!
World models for evaluating autonomous driving, GAIA3 released! End-to-end driving model &loads of insights!
Thanks for visiting &spending the day talking to researchers.
World models for evaluating autonomous driving, GAIA3 released! End-to-end driving model &loads of insights!
Thanks for visiting &spending the day talking to researchers.
December 2, 2025 at 3:26 PM
Super-exciting talk by Ani Kembhavi from Wayve AI @bristoluni.bsky.social @compscibristol.bsky.social #MaVi Seminar today!
World models for evaluating autonomous driving, GAIA3 released! End-to-end driving model &loads of insights!
Thanks for visiting &spending the day talking to researchers.
World models for evaluating autonomous driving, GAIA3 released! End-to-end driving model &loads of insights!
Thanks for visiting &spending the day talking to researchers.
Reposted by Evangelos Kazakos
Going to #NeurIPS2025 San Diego? Escape the conference for a couple hours with a morning bird walk in the trails of Balboa Park. 7 am on Thurs, Dec. 4.
November 28, 2025 at 2:14 PM
Going to #NeurIPS2025 San Diego? Escape the conference for a couple hours with a morning bird walk in the trails of Balboa Park. 7 am on Thurs, Dec. 4.
Reposted by Evangelos Kazakos
🧠 How do neurons encode information? We know HOW MUCH, but what about WHAT information they encode?
Our new work uses diffusion models to decompose neural information down to individual stimuli & features!
🎯Spotlight at #NeurIPS2025 🌟📄
arxiv.org/abs/2505.11309
Our new work uses diffusion models to decompose neural information down to individual stimuli & features!
🎯Spotlight at #NeurIPS2025 🌟📄
arxiv.org/abs/2505.11309
December 1, 2025 at 1:12 PM
🧠 How do neurons encode information? We know HOW MUCH, but what about WHAT information they encode?
Our new work uses diffusion models to decompose neural information down to individual stimuli & features!
🎯Spotlight at #NeurIPS2025 🌟📄
arxiv.org/abs/2505.11309
Our new work uses diffusion models to decompose neural information down to individual stimuli & features!
🎯Spotlight at #NeurIPS2025 🌟📄
arxiv.org/abs/2505.11309
Reposted by Evangelos Kazakos
🧠🔬 Excited to share our #NeurIPS2025 paper: "Convolution Goes Higher-Order"!
We asked: Can shallow networks be as expressive as deep ones? Inspired by biological vision, we introduce higher-order convolutions that capture complex image patterns standard CNNs miss.
🧵👇
We asked: Can shallow networks be as expressive as deep ones? Inspired by biological vision, we introduce higher-order convolutions that capture complex image patterns standard CNNs miss.
🧵👇
December 1, 2025 at 1:24 PM
🧠🔬 Excited to share our #NeurIPS2025 paper: "Convolution Goes Higher-Order"!
We asked: Can shallow networks be as expressive as deep ones? Inspired by biological vision, we introduce higher-order convolutions that capture complex image patterns standard CNNs miss.
🧵👇
We asked: Can shallow networks be as expressive as deep ones? Inspired by biological vision, we introduce higher-order convolutions that capture complex image patterns standard CNNs miss.
🧵👇
Reposted by Evangelos Kazakos
A NeurIPS feed! Pin it and follow along.
December 1, 2025 at 1:49 PM
A NeurIPS feed! Pin it and follow along.
First time that I'm attempting to read a paper with symbolic reasoning.
Am I confused? Yes!
Am I hyped? Also yesss!
Congrats @toniwuest.bsky.social et al. for a great paper and for convincing me to look into symbolic reasoning. It makes a lot of sense, especially within VLMs!
Am I confused? Yes!
Am I hyped? Also yesss!
Congrats @toniwuest.bsky.social et al. for a great paper and for convincing me to look into symbolic reasoning. It makes a lot of sense, especially within VLMs!
🚨 New paper alert!
We introduce Vision-Language Programs (VLP), a neuro-symbolic framework that combines the perceptual power of VLMs with program synthesis for robust visual reasoning.
We introduce Vision-Language Programs (VLP), a neuro-symbolic framework that combines the perceptual power of VLMs with program synthesis for robust visual reasoning.
December 1, 2025 at 3:48 PM
First time that I'm attempting to read a paper with symbolic reasoning.
Am I confused? Yes!
Am I hyped? Also yesss!
Congrats @toniwuest.bsky.social et al. for a great paper and for convincing me to look into symbolic reasoning. It makes a lot of sense, especially within VLMs!
Am I confused? Yes!
Am I hyped? Also yesss!
Congrats @toniwuest.bsky.social et al. for a great paper and for convincing me to look into symbolic reasoning. It makes a lot of sense, especially within VLMs!
Reposted by Evangelos Kazakos
🚨 New paper alert!
We introduce Vision-Language Programs (VLP), a neuro-symbolic framework that combines the perceptual power of VLMs with program synthesis for robust visual reasoning.
We introduce Vision-Language Programs (VLP), a neuro-symbolic framework that combines the perceptual power of VLMs with program synthesis for robust visual reasoning.
November 30, 2025 at 1:32 AM
🚨 New paper alert!
We introduce Vision-Language Programs (VLP), a neuro-symbolic framework that combines the perceptual power of VLMs with program synthesis for robust visual reasoning.
We introduce Vision-Language Programs (VLP), a neuro-symbolic framework that combines the perceptual power of VLMs with program synthesis for robust visual reasoning.
Reposted by Evangelos Kazakos
I'm open to there being a role for blind review, but introducing non-blind review has a lot of upsides that may reduce how much we actually care about blind review.
I think we care about blind review only because our publishing system is poorly designed and needs change in the modern era anyway.
I think we care about blind review only because our publishing system is poorly designed and needs change in the modern era anyway.
The only argument advanced by proponents of blind peer-review boils down to "less powerful people can't criticize powerful people in public," the same argument people make when advocating for anonymity on social media.
In light of the new OpenReview identity-leak scandal, it's a good time to question our assumptions about why blinding in peer review is helpful in the first place.
November 30, 2025 at 4:13 PM
I'm open to there being a role for blind review, but introducing non-blind review has a lot of upsides that may reduce how much we actually care about blind review.
I think we care about blind review only because our publishing system is poorly designed and needs change in the modern era anyway.
I think we care about blind review only because our publishing system is poorly designed and needs change in the modern era anyway.
Weeding Dub Gypsy Dub feat Ras Divarius
YouTube video by Skeuf Saint
youtu.be
December 1, 2025 at 6:48 AM
Reposted by Evangelos Kazakos
And a major open science release from Prime Intellect: they don’t stress it enough but SFT part is beyond post-training. This is a fully documented mid-training with tons of insights/gems on MoE training, asynchronous infra RL, deep research. storage.googleapis.com/intellect-3-...
November 27, 2025 at 7:47 AM
And a major open science release from Prime Intellect: they don’t stress it enough but SFT part is beyond post-training. This is a fully documented mid-training with tons of insights/gems on MoE training, asynchronous infra RL, deep research. storage.googleapis.com/intellect-3-...
Reposted by Evangelos Kazakos
Out of curiosity I decided to try and run the numbers on how much Netflix you can watch for the energy cost of a ChatGPT prompt
As far as I can tell it's between 5.1 and 10.2 seconds, depending on which end of the 2019 IEA Netflix energy usage estimate you use
simonwillison.net/2025/Nov/29/...
As far as I can tell it's between 5.1 and 10.2 seconds, depending on which end of the 2019 IEA Netflix energy usage estimate you use
simonwillison.net/2025/Nov/29/...
November 29, 2025 at 2:16 AM
Out of curiosity I decided to try and run the numbers on how much Netflix you can watch for the energy cost of a ChatGPT prompt
As far as I can tell it's between 5.1 and 10.2 seconds, depending on which end of the 2019 IEA Netflix energy usage estimate you use
simonwillison.net/2025/Nov/29/...
As far as I can tell it's between 5.1 and 10.2 seconds, depending on which end of the 2019 IEA Netflix energy usage estimate you use
simonwillison.net/2025/Nov/29/...
Reposted by Evangelos Kazakos
In light of the new OpenReview identity-leak scandal, it's a good time to question our assumptions about why blinding in peer review is helpful in the first place.
November 29, 2025 at 10:08 AM
In light of the new OpenReview identity-leak scandal, it's a good time to question our assumptions about why blinding in peer review is helpful in the first place.
Reposted by Evangelos Kazakos
she shared excellent insights on what this implies (your brain runs probably faster on those questions, when you a future CVPR PC 😅):
- Nothing will ever be blind anymore because there's enough data in the wild to train a reviewer de-anonymizer
- What about fine-tuning a seq2seq model conditioned
- Nothing will ever be blind anymore because there's enough data in the wild to train a reviewer de-anonymizer
- What about fine-tuning a seq2seq model conditioned
November 28, 2025 at 4:42 PM
she shared excellent insights on what this implies (your brain runs probably faster on those questions, when you a future CVPR PC 😅):
- Nothing will ever be blind anymore because there's enough data in the wild to train a reviewer de-anonymizer
- What about fine-tuning a seq2seq model conditioned
- Nothing will ever be blind anymore because there's enough data in the wild to train a reviewer de-anonymizer
- What about fine-tuning a seq2seq model conditioned
Fully agree with Victor on this one.
💯 crucifying @openreview.bsky.social (or even @iclr-conf.bsky.social ) for the incident is exaggerated.
100% Reliability doesn't exist! Ask Boeing engineers
Btw, I understand the anger & fear. Now, pls disclose other similar stands of yours, eg measures & actions after Boeing ✈️ incidents.
100% Reliability doesn't exist! Ask Boeing engineers
Btw, I understand the anger & fear. Now, pls disclose other similar stands of yours, eg measures & actions after Boeing ✈️ incidents.
I think they try their best, but it’s not a trivial business… I can actually understand arguments for both sides, but OR integration has so signifincantly progressed, I don’t think there will be a going back … would still like OR to fix some more issues thought… 🫠
November 29, 2025 at 10:27 AM
Fully agree with Victor on this one.
Reposted by Evangelos Kazakos
I got this per Whatsapp, don't know who created it:
November 28, 2025 at 5:31 PM
I got this per Whatsapp, don't know who created it:
Reposted by Evangelos Kazakos
really annoying.
my student just made this joke:
ICLR= I Can Locate Reviewers 😅
my student just made this joke:
ICLR= I Can Locate Reviewers 😅
November 28, 2025 at 3:41 PM
really annoying.
my student just made this joke:
ICLR= I Can Locate Reviewers 😅
my student just made this joke:
ICLR= I Can Locate Reviewers 😅
Reposted by Evangelos Kazakos
#KostasThoughts: The #ICLR2026 review process has been disappointing. This is not on the PCs. It reflects our current environment where the scale of reviewing across the various conferences may have outgrown the bandwidth of the community (eg shortcuts via LLM generated reviews) …
1/2
1/2
Statement from #ICLR2026 Program Chairs
November 28, 2025 at 4:32 PM
#KostasThoughts: The #ICLR2026 review process has been disappointing. This is not on the PCs. It reflects our current environment where the scale of reviewing across the various conferences may have outgrown the bandwidth of the community (eg shortcuts via LLM generated reviews) …
1/2
1/2