Depth Anything 3 Bytedance Seed proposes a simple & scalable unified model for recovering 3D visual space from any number of views with or without known camera poses. Absolute top notch quality with SoTA 3D reconstruction and camera pose estimation
MACHINE LEARNING
-
AlphaFold 2 celebrates five years transforming global research
By
–
Thrilled to celebrate 5 years of AlphaFold 2! It’s now been used by over 3 million researchers around the world to accelerate their vital research – and it was an honour of a lifetime for our work to be recognised last year with the Nobel Prize! Proof of AI’s potential to enable
-
AlphaFold Five Years Impact and The Thinking Game Documentary
By
–
Read more about AlphaFold in our blog: https://
deepmind.google/blog/alphafold
-five-years-of-impact/
… Watch The Thinking Game on YouTube (a sort of sequel to the award-winning AlphaGo documentary – if you enjoyed that, you will enjoy this): -
ROOT: Robust Orthogonalized Optimizer for Neural Network Training
By
–
ROOT: Robust Orthogonalized Optimizer for Neural Network Training Huawei Noah's Ark Lab
Paper: https://
arxiv.org/abs/2511.20626
Code: https://
github.com/huawei-noah/no
ah-research/tree/master/ROOT
… Our report: https://
mp.weixin.qq.com/s/X7dNh8lwr0xV
W7TsuO4D2g
… -
Huawei ROOT: Robust Orthogonalized Optimizer for LLM Training
By
–
This new optimizer can make training giant LLMs both more stable and more precise, even under noise and extreme scale! Huawei just introduces ROOT, a Robust Orthogonalized Optimizer that tackles two big weaknesses in recent momentum-orthogonalized methods: – Dimensional
-
Faster R-CNN Wins NeurIPS Test of Time Paper Award
By
–
Congratulations to Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun! Their paper “Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks” has been awarded the Test of Time Paper Award at NeurIPS 2025. The Faster R-CNN paper has been cited more than
-
ReSeek: Self-Correcting Search Agents with Dense Rewards
By
–
Search agents could notice their own mistakes and fix their reasoning mid-trajectory. ReSeek enables exactly that. With a JUDGE action for on the fly self correction and dense rewards for factual correctness plus real utility, it trains agents that outperform SOTA on complex
-
Self-Critique in Agentic AI: Benefits and Risks Explained
By
–
Self-critique is a core pattern in agentic AI — but it turns out to be dangerous in the wrong place.
New blog from @ArminPCM show: Self-critique helps when models are failing It destroys accuracy when models are already right
This has big implications for RLAIF, RLFT, and -
Self-Verification as Adversarial Training Signal for AI Models
By
–
The key takeaway:
Critique is for debugging — not polishing.
When the same model plays both “actor” and “judge,” self-verification can become an adversarial training signal.
Read the full breakdown → -
ScaleNet: Scaling Vision Transformers with Weight Sharing
By
–
What if you could scale up a ViT in minutes instead of training a giant one from scratch? ScaleNet makes it possible by inserting weight shared layers into pretrained ViTs with tiny adapter adjustments. It expands depth with almost no extra parameters, boosts a 2x depth DeiT