Reinforcement learning for LLMs—fully open-sourced DAPO trains Qwen2.5-32B with RL, hitting 50 points on AIME 2024, outperforming DeepSeek-R1-Zero after just 50% of the training steps. Open-source code & dataset
Improved training stability Trending #1 on alphaXiv
RESEARCH
-

DAPO: Open-Source RL Training for Qwen2.5-32B LLMs
By
–
-
Stanford Researchers Use LLMs to Personalize Middle School Lessons
By
–
The learning gap in middle school classrooms has widened due to the pandemic. Discover how @Stanford
's education and computer science researchers are leveraging large language models to help teachers create lessons that meet every student's needs: -
Hugging Face 2024: Datasets and Tools Outpace Models Usage
By
–
For the first time in 2024 on @huggingface
:
– usage on datasets/spaces/papers > usage on models
– usage on all other domains > usage on text
– usage from AI builders (software engineers,…) > usage from AI scientists Slowly but surely becoming the platform that all AI builders -
AI Models Cannot Learn During Interactions: Debate Impact Analysis
By
–
The cognitive dissonance here is that these models don't learn during interactions. No online learning: it's literally the most fruitless debate you can imagine for its participants. The debate only has real impact in the minds of humans reading it, if they exist. https://
x.com/LiorOnAI/statu
/LiorOnAI/status/1903017981632020682
… -

Asynchronous RL Agents Learn Faster, Achieve Higher Returns
By
–
A not-so-surprising finding: with asynchronous RL, agents learn faster and achieve significantly higher returns. Still, it's nice to see it empirically verified.
-
Stanford’s Gaussian Atlas: 3D Scene Representation Breakthrough
By
–
Project Page: https://
cs.stanford.edu/~xtiange/proje
cts/gaussianatlas/
…
Paper: https://
arxiv.org/abs/2503.15877 -

GaussianVerse: Large-Scale 3D Gaussian Dataset Released
By
–
They also built GaussianVerse, a large-scale dataset with 205K high-quality 3D Gaussian fittings of various 3D objects.
-
2D Diffusion Models Repurposed for 3D Object Generation
By
–
A team led by Dr. Li Fei-Fei @drfeifei found that pre-trained 2D diffusion models can be repurposed for 3D object generation. Their proposed GaussianAtlas enables fine-tuning of 2D diffusion models to generate 3D Gaussians. pic.twitter.com/D2mvz52C08
— 机器之心 JIQIZHIXIN (@jiqizhixin) 21 mars 2025A team led by Dr. Li Fei-Fei @drfeifei found that pre-trained 2D diffusion models can be repurposed for 3D object generation. Their proposed GaussianAtlas enables fine-tuning of 2D diffusion models to generate 3D Gaussians.
-
US talent concentration in AI: fact or misconception?
By
–
Most talent is in the US!? what do you base this on?
-

X-Attention: Novel Mechanism for Efficient AI Model Processing
By
–
Paper: https://
arxiv.org/pdf/2503.16428
Code: https://
github.com/mit-han-lab/x-
attention
…