Bora Kargı

💻 ML Msc. @ University of Tübingen || METU CENG'23
💻 Research Engineer @ ELLIS Institute Tübingen (OpenEuroLLM)
opinions are my own
Opinio

[ELLIS Institute Tübingen @ELLISInst_Tue]

New funding for AI safety research! 🎉
ELLIS Institute Tübingen PI @sahar_abdelnabi has received funding from @coeff_giving to support research on developing context-sensitive benchmarks and training methods for sycophancy in AI systems.
This work will help advance the development of safe, aligned, and reliable AI.
Congratulations, Sahar! 👏

[Bora Kargı @bora_kargi]

@tomhacks @Sentdex There are some attempts to make it everything fully open like OLMO3 (https://t.co/X6aw7sNTID). I feel like it would be nice to have this type of releases for other scales/models as well, but possibly due to many other problematic samples they use, it will be not possible..

[Gege Gao @GegeHeseri]

If you were asked to paint this picture, how would you do it? (1/n)
Most humans would not paint every pixel at once.
This simple intuition inspired our ECCV 2026 paper: Trajectory Forcing.

Project page: https://t.co/PbKMY5XTBP

#ECCV2026

[Jonas Geiping @jonasgeiping]

We’re training models wrong and it’s due to chatGPT. Even the modern coding agents used daily still use message-based exchanges: They send messages to users, to themselves (CoT) and to tools, and receive messages in turn.

This bottlenecks even very intelligent agents to a single stream. The models cannot read while writing, cannot act while thinking and cannot think while processing information.

In our new paper, see below, we discuss LLMs with parallel streams. We show that multi-stream LLMs can …
🔵Be created by instruction-tuning for the stream format

🔵Simplify user and tool use UX removing many pain points with agents and chat models (such as having to interrupt the model to get a word in)

🔵Multi-Stream LLMs are fast, they can predict+read tokens in all streams in parallel in each forward pass, improving latency

🔵 LLMs with multiple streams have an easier time encoding a separation of concerns, improving security

🔵 LLMs with many internal streams provide a legible form of parallel/cont. reasoning. Even if the main CoT stream is accidentally pressured or too focused on a particular task to voice concerns, other internal streams can subvocalize concerns that would otherwise not be verbalized.

Does this sound related to a recent thinky post :) - Yes, but I don’t feel so bad about being outshipped with such a cool report on their side by 23 hours. I’ll link a 2nd thread below with a more direct comparison. I actually think both are complementary in interesting ways.

[ELLIS Institute Tübingen @ELLISInst_Tue]

From research to unicorn in just 18 months!

What a milestone for our PI @FrankRHutter and his team at @prior_labs .
The whole team at the ELLIS Institute Tübingen is proud of you all.
What a motivation for all of us going forward 🚀

[Arnas Uselis @a_uselis]

How do embedding spaces of models that generalize from limited data look?

We study what structure such models should exhibit.

Turns out: linear and orthogonal. And modern embedding models like CLIP and SigLIP already show signs of it!

🧵 (1/n)

[Bora Kargı @bora_kargi]

RT @d_yesiltepe: Infinity-RoPE is accepted to #CVPR2026🎉!

✨In Infinity-RoPE, we extend pretrained autoregressive video models beyond the t…

[Solcu Gazete @solcugazete60]

Eylemden sonra dağılan kadınlar Taksim Tünel'de polis ablukasına alındı.

[elvis @omarsar0]

Leaky Thoughts

Hey AI devs, be careful how you prompt reasoning models.

This work shows that reasoning traces frequently contain sensitive user data.

More of my notes below:

[Bora Kargı @bora_kargi]

@gowerrobert people be forgetting about Taylor Approximation 🫡

[Bloomberg @business]

Istanbul’s mayor Ekrem Imamoglu was detained at his home Wednesday morning, CNNTurk reports, in a move that could bar him from challenging Recep Tayyip Erdogan in the next presidential election https://t.co/HlJN9iZqKF

[Bora Kargı @bora_kargi]

@Darkwebhaber 53bin kisi ne demek ya (ki daha fazla)
Bu ulkenin balik hafizasi ne zaman duzelcek? Bir sehir öldü ve kimse cezasini cekmedi??

[Matthias Niessner @MattNiessner]

📢📢 𝐏𝐫𝐄𝐝𝐢𝐭𝐨𝐫𝟑𝐃: 𝐅𝐚𝐬𝐭 𝐚𝐧𝐝 𝐏𝐫𝐞𝐜𝐢𝐬𝐞 𝟑𝐃 𝐒𝐡𝐚𝐩𝐞 𝐄𝐝𝐢𝐭𝐢𝐧𝐠 📢📢

We propose a training-free 3D shape editing approach that rapidly and precisely edits the regions intended by the user and keeps the rest as is.

Using a quickly brushed mask and a text prompt, we first apply multi-view editing in the 2D domain and then run our merging algorithm in the 3D feature space to ensure that the edited shape is loyal to the input shape.

Project Page: https://t.co/QRRcF1AP7Q
Video: https://t.co/pMOsbpYUKf

Great work by @ErkocZiya @cangumeli Chaoyang Wang @angelaqdai @peter_wonka @hyjameslee @PeiyeZ

[Hidir Yesiltepe @d_yesiltepe]

🚀 Excited to share our latest work, MotionShop, a training-free approach for motion transfer in video diffusion models! 🎥 Big thanks to my amazing teammates from GEMLAB—@tunahansalih, Connor Dunlop, and @PINguAR—for making this possible! 🙌

🌐 https://t.co/6cP67aFdSV

[Bora Kargı @bora_kargi]

@haskologlu Ülkenin geldiği duruma bak Yemeksepeti CEO'su insanların annesine sövüyor

[Bora Kargı @bora_kargi]

@andreasgeiger0 Very excited for this!

[Seong Joon Oh @coallaoh]

Excited to share a preprint at STAI! It's a study on the user perspective in training data attribution, but my secret, meta-level research question behind the work is: Can AI research also be user-driven?

As researchers, we often theorise about user needs, but we tend to overlook the importance of direct engagement with the actual users. By speaking with them, we’ve gained insights into how AI technologies must evolve to meet actual demands. They were quite different from my initial assumptions.

So far, especially after this work, my conclusion is - yes, AI research can and should be user-driven.

Authors: @_elinguyen, Johannes Bertram, @EKortukov, @jeanysong

[adil meric @adilmeric12]

Introducing the first NeRF-based 3D style transfer method that generalizes across scenes and styles!

Our approach eliminates the time-consuming optimization for each scene or style, making it more efficient than previous NeRF-based methods. Check it out! #GCPR2024

[Matthias Niessner @MattNiessner]

G3DST: Generalizing 3D Style Transfer with NeRF across Scenes and Styles!

Given a style latent, our hypernetwork estimates MLP params that transform aggregated ray features.

https://t.co/6pOxGoZMuG

Great work by our MA student @adilmeric12
U. Kocasari @barbara_roessle #GCPR24