In RLC 2024 @RL_conference, we shall present a recent work about robust reinforcement learning to handle linear function approximation and robustness together in offline settings: “Sample Complexity of Offline Distributionally Robust Linear Markov Decision Processes”

#MLSys2025 call for papers is out! The conference will be led by the general chair Matei Zaharia , PC chairs Yingyan (Celine) Lin, and Gauri Joshi. Consider submitting and bringing your latest works in AI and systems—more details at mlsys.org.

Please check out this thread where my amazing student Harry Dong explains his exciting work to be presented at COLM on efficient LLM inference! Joint work with equally amazing colleague Beidi Chen

Theory papers should not perform experiments, because once they do, they’ll be criticized for having too much theory that “reduced the volume of the experimental contents”.

I am not at #neurips but just heard about this incident and feel very upset about it. This specific calling out of the country of origin is offensive, inappropriate and completely unnecessary. This also breaks the code of conduct and @neurips should take her talk off the website.

Thank you for the quick reaction to this matter! It means a lot. I hope you’ll remove her talk from the website.

I will bet most of these high achieving students come from privileged labs/universities. In admission, we should calibrate these factors and look for potentials and out-of-box thinkers. (Opinions are my own.)

Amazing list of resources for early-career researchers in ML/data science/AI put together by my amazier student Laixi, who happens to be on the market this year!

Another great edition of #MLSys is on the way!

Some of us who work in higher ed but not close enough to AI need the alarm call to wake up and move faster. We need to rethink our curriculum and pedagogical approaches!

Today is the start of a new era of natively multimodal AI innovation. Today, we’re introducing the first Llama 4 models: Llama 4 Scout and Llama 4 Maverick — our most advanced models yet and the best in their class for multimodality. Llama 4 Scout • 17B-active-parameter model

d1: to grow in reasoning, masked diffusion language models go beyond supervised learning , we meet RL !😃dllm-reasoning.github.io

🧠 Your LLM should model how you think, not reduce you to preassigned traits 📢 Introducing LoRe: a low-rank reward modeling framework for personalized RLHF ❌ Demographic grouping/handcrafted traits ✅ Infers implicit preferences ✅ Few-shot adaptation 📄 arxiv.org/abs/2504.14439

Serious issues with AC paper bidding NeurIPS Conference Paper matching ignored my excluded papers, gave me mostly papers on topics I have little to no expertise in and we can only bid on limited subset of these mostly irrelevant papers. Friend reported having the same issue #NeurIPS25

Why do gradients increase near the end of training? Read the paper to find out! We also propose a simple fix to AdamW that keeps gradient norms better behaved throughout training. arxiv.org/abs/2506.02285

🚨 Code is live! Check out LoRe – a modular, lightweight codebase for personalized reward modeling from user preferences. 📦 Few-shot personalization 📊 Benchmarks: TLDR, PRISM, PersonalLLM 👉 github.com/facebookresear… Huge thanks to AI at Meta for open-sourcing this research 🙌

📢Exciting updates from #MLSys2025! All session recordings are now available and free to watch at mlsys.org. We’re also thrilled to announce that #MLSys2026 will be held in Seattle next May—submissions open next month with a deadline of Oct 30. We look forward to

Looking forward to this exciting workshop!

🚨 🔥 Multi-step reasoning is key to solving complex problems — and Transformers with Chain-of-Thought can do it surprisingly well. 🤔 But how does CoT function as a learned scratchpad that lets even shallow Transformers run sequential algorithms that would otherwise require

A one-layer multi-head transformer, with CoT, enables both forward and reversal reasoning. The training dynamics analysis particularly illuminates how two (heads) are better than one! See Tong’s post below. Joint work with Tong Yang Yu Huang and Yingbin Liang.