🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

Please come to our poster today — I’ll be there to present our work! “A Coefficient Makes SVRG Effective” Friday, 3:00–5:30 Hall 3 + Hall 2B #385 iclr.cc/virtual/2025/p…

our ICLR 2025 work on making a classic optimization technique practically effective in deep learning training I'm by no means an optimization expert - credits to my incoming PhD student David Yin for leading this work

Not at ICLR myself, but David Yin will kindly give the oral and poster presentation on my behalf, for our dataset bias paper. David knows the work as well as I know it (if not better). Check it out, happening today! Oral: Saturday, 10:42 — 10:54am, Peridot 202-203

👉👉#ICLR2025 #SCOPE workshop has just kicked off, now!!!! Those who are in person, please come and join us at Singapore expo, Peridot 204-205 ( 9am - 6pm Singapore time). Schedule here: scope-workshop.github.io Keynote talks in morning session (9 am onwards): Yu Cheng

Check out our new paper on training multi-modal LLMs to answer complex visual questions, lead by superstar students Xindi Wu Will Hwang ! We show that increasing the representation of compositionally complex questions in visual instruction tuning data boosts performance

Accepted to #ICML 25 & also recently featured in CMU news and Fast Company: cs.cmu.edu/news/2025/llm-… fastcompany.com/91286162/ai-ch…

COLM reviewer guideline is next level. I am touched and cannot agree more. Conference on Language Modeling

Still relying on OpenAI’s CLIP — a model released 4 years ago with limited architecture configurations — for your Multimodal LLMs? 🚧 We’re excited to announce OpenVision: a fully open, cost-effective family of advanced vision encoders that match or surpass OpenAI’s CLIP and

Excited to share our new method ✏️PENCIL! It decouples space complexity from time complexity in LLM reasoning, by allowing model to recursively erase and generate thoughts. Joint work w. my student Chenxiao Yang , along with Nati Srebro Bartom and David McAllester.

Very surprised at what a non-reasoning model (4o) can do. The first answer, "Long Branch, NJ", is correct It's a new chat session with memory turned off. Not sure how it arrived at it

And this one I asked it why. It did provide some valid reasoning, but not sure how these together leads to "strongly resembles Long Branch, New Jersey" new session still in 4o, not a "reasoning" model per se

This is almost exactly what I've been working on / thinking about lately

Angjoo Kanazawa Angjoo Kanazawa and I taught CS 280, graduate computer vision, this semester at UC Berkeley. We found a combination of classical and modern CV material that worked well, and are happy to share our lecture material from the class. cs280-berkeley.github.io Enjoy!

Big congrats to TaiMing Lu on the research award & CRA finalist! And shoutout to Daniel Khashabi 🕊️ for the teaching award—NLP:SSM (self-supervised.cs.jhu.edu) is a must-take!

They study model merging (EMA, soups), which is well-understood for fine-tunings of a base model, but they investigate *pre-training* of LLMs. My summary thread. (YoU wOn'T bEliEvE the surprise in post #5 which makes me like this group a lot more!)