Nebius Science

A Nebius perspective on AI-driven research.

[Nebius Science @nebius_science]

Why we picked these specific papers? Check out the full article: https://t.co/qFvUA4Jn1A

(5/5)

[Nebius Science @nebius_science]

✔️ A frontier model that puts our own speculative decoding research to work
✔️ An RL recipe for fixing retrieval without touching the retriever
✔️ Evidence that vision-language models stop looking when you ask the wrong question

Check out the papers we're reading!

(1/5)

[Nebius Science @nebius_science]

Tinted Frames: Question Framing Blinds Vision-Language Models. Ask open-ended questions when you can, because the model literally looks harder: https://t.co/avS6upRRlX

(4/5)

[Nebius Science @nebius_science]

Nebius Science Paper Club is a webinar series led by Nebius researchers, bringing together paper authors and practitioners to discuss new ideas and discoveries in AI.

The paper that we're diving into on August 25: https://t.co/AEa5cSyib5

[Nebius Science @nebius_science]

Join @federicobianchy, Staff Scientist at @togethercompute, for our next Paper Club session, focused on Learning to Discover at Test Time (TTT-Discover).

📆 August 25, 2026
⏰ 6:00 PM CEST
📹 Online

Register now: https://t.co/Qlzjc4grMm

[Nebius Science @nebius_science]

Dive into details and learn more about why we picked these specific papers: https://t.co/VI5lsoYnBL

(5/5)

[Nebius Science @nebius_science]

Kimi K3, and two papers our team argued about this week. The Nebius R&D team digs into papers so you don't have to.

(1/5)

[Nebius Science @nebius_science]

Stable Asynchrony: Variance-Controlled Off-Policy RL for LLMs. When rollouts are stale, a few unrepresentative samples can dominate each update. The authors show you can see the blow-up coming by tracking the effective sample size: https://t.co/D5zERdpP5M

(4/5)

[Nebius Science @nebius_science]

8/ Full write-up with why we picked each one: https://t.co/3bUuAC8ccl

What should we be reading next week? Replies open 👇

[Nebius Science @nebius_science]

1/ Paper Picks #2 is out 🧵

Three papers that caught our R&D teams' attention this week: production-grade speculative decoding, proposal trees from factorized priors, and attention without sinks.

[Nebius Science @nebius_science]

7/ Sigmoid attention instead of softmax + RoPE with balanced ALiBi = information compressed and retained within a plain sliding window. Matches or beats SOTA linear recurrent architectures on 8 benchmarks.