Q-learning is not yet scalable seohong.me/blog/q-learnin… I wrote a blog post about my thoughts on scalable RL algorithms. To be clear, I'm still highly optimistic about off-policy RL and Q-learning! I just think we haven't found the right solution yet (the post discusses why).

Sparsity can make your LoRA fine-tuning go brrr 💨 Announcing SparseLoRA (ICML 2025): up to 1.6-1.9x faster LLM fine-tuning (2.2x less FLOPs) via contextual sparsity, while maintaining performance on tasks like math, coding, chat, and ARC-AGI 🤯 🧵1/ z-lab.ai/projects/spars…

See you in Hawaii🖖🏝️

I converted one of my favorite talks I've given over the past year into a blog post. "On the Tradeoffs of SSMs and Transformers" (or: tokens are bullshit) In a few days, we'll release what I believe is the next major advance for architectures.

Everyone knows action chunking is great for imitation learning. It turns out that we can extend its success to RL to better leverage prior data for improved exploration and online sample efficiency! colinqiyangli.github.io/qc/ The recipe to achieve this is incredibly simple. 🧵 1/N

For everyone interested in precise 📷camera control 📷 in transformers [e.g., video / world model etc] Stop settling for Plücker raymaps -- use camera-aware relative PE in your attention layers, like RoPE (for LLMs) but for cameras! Paper & code: liruilong.cn/prope/

Check out our blog post at flowreinforce.github.io  We developed interactive plots that explain the connection between flow/diffusion models and RL. w/ a great team of collaborators! Songwei Ge Brent Yi Chung Min Kim Ethan Weber Hongsuk Benjamin Choi Haiwen (Haven) Feng Angjoo Kanazawa

Viser is a gift to 3D/4D community!!!

Genie3 is so incredible!

Glad to see gsplat powering real-time physics simulation together with NVIDIA Warp! Learn more 👉 nvda.ws/4m4LGnN

📢Thrilled to share that I'll be joining Harvard and the Kempner Institute as an Assistant Professor starting Fall 2026! I'll be recruiting students this year for the Fall 2026 admissions cycle. Hope you apply!

Introducing 𝗦𝗧𝗿𝗲𝗮𝗺𝟯𝗥, a new 3D geometric foundation model for efficient 3D reconstruction from streaming input. Similar to LLMs, STream3R uses casual attention during training and KVCache at inference. No need to worry about post-alignment or reconstructing from scratch.

Excited to see Orthogonal Finetuning (OFT) and Quantized OFT (QOFT) now merged into LLaMA-Factory! 🎉 OFT & QOFT are memory/time/parameter-efficient and excel at preserving pretraining knowledge. Try them in: 🔗 LLaMA-Factory: github.com/hiyouga/LLaMA-… 🔗 PEFT:

Introducing Phota Labs: We're building personalized visual GenAI and the next chapter of photography. Because memory-making should be effortless, personal, and compelling for everyone. We're excited to share our $5.6M seed led by a16z ( Yoko ), with @Figma Ventures,

We'd like thank reviewers and community that 4DGT got accepted to NeurIPS 2025 as a Spotlight. We have just released the demo code in github.com/facebookresear… There are a few features to be added with some updates in our writing, thanks to the awesome suggestions from

Thrilled to announced that at #ICCV2025 we will host the first workshop on 𝐆𝐞𝐨𝐦𝐞𝐭𝐫𝐲-𝐅𝐫𝐞𝐞 𝐍𝐨𝐯𝐞𝐥 𝐕𝐢𝐞𝐰 𝐒𝐲𝐧𝐭𝐡𝐞𝐬𝐢𝐬 𝐚𝐧𝐝 𝐂𝐨𝐧𝐭𝐫𝐨𝐥𝐥𝐚𝐛𝐥𝐞 𝐕𝐢𝐝𝐞𝐨 𝐌𝐨𝐝𝐞𝐥𝐬 geofreenvs.github.io a.k.a. "3D Computer Vision in the era of Video Models" 😅

📢 Lyra: Generative 3D Scene Reconstruction via Video Diffusion Model Self-Distillation Got only one or a few images and wondering if recovering the 3D environment is a reconstruction or generation problem? Why not do it with a generative reconstruction model! We show that a

I enjoy reading this blog. This is exactly what I am trying to pursue throughout my research career -- using weight geometry to characterize and improve neural network training. Really excited that it finally got people's attention now!! In 2017, we study the weight

🚨 New preprint. Physics of learning: A Lagrangian perspective to different learning paradigms. arxiv.org/abs/2509.21049 TL;DR A single Lagrangian unifies supervised, generative modelling, and RL. - We study the problem of building an efficient learning system. - We propose that