Professor at Texas A&M University; ML/AI researcher; optimization for ML/AI; large reasoning models, developing LibAUC library for training deep neural nets.
College Station, TX
Joined March 2016
- Tweets219
- Following1.1K
- Followers675
- Likes607
Impressive results! Strongly agree that compute efficiency is a critical frontier in AI. Our DRRHO-CLIP outperforms OpenAI’s CLIP with 17× less compute. The future of AI is not just about more GPUs, but more intelligence per GPU. Great to see Physion pushing in this direction!
The AI industry has a massive compute efficiency problem. At Physion Labs, we think it's time to challenge it.
Today, we're sharing the technology behind NexTron, our world-intelligence training engine.
Our benchmarks on matched hardware:
⚡ Up to 20.9× SFT throughput vs. DeepSpeed (H100)
⚡ Up to 6.6× SFT throughput vs. Megatron (H200)
⚡ Up to 7.2× SFT throughput vs. Megatron (B200)
⚡ Up to 3.1× end-to-end RL throughput vs. Miles–Megatron
🔥 180B model fine-tuning on just 8 H100 GPUs.
Same hardware. Dramatically higher throughput. Comparable model quality.
We believe the next frontier isn't just building bigger GPU clusters.
It's unlocking more intelligence from every GPU, every watt, and every dollar spent on training.
#PhysionLabs #NexTron #WorldIntelligence #AIInfrastructure
Tianbao (TB) Yang retweeted
1/13. Rufus-Air (arxiv.org/abs/2609.29421) 🧵
Sep 2025–Apr 2026, minus a 2-month production pause: a small research team at Amazon Search built an open post-training recipe from scratch on GLM-4.5-Air-Base, using public data and open-source components, rarely over 400 H200s.
ChatGPT: “We’re giving researchers free access.” Great—let’s verify my researcher status. Failed. Again and again.
Claude: “We’re giving scientists free access.” Great—here’s my lab website. Not verified.
And we’re talking about AGI? 😅
Those who use ChatGPT to solve challenging math problems and write a paper saying so may undermine theoretical research, discouraging students and prompting reviewers to question researchers’ original contributions.
3/n I have worked on this problem for many years but struggle to design an elegant algorithm which enjoy fast convergence both theoretically and practically.
3. SCENT adopts mirror descent but differs in that the distance generating function exp(-x) is never used before.
1/n Today, I will present the most interesting algorithm that I have worked on in my career so far.
- solves a fundamental objective in ML, i.e. log-expectation-exponential (entropic risk).
- a novel synthesis of mirror descent and stochastic proximal point methods.
A new learning theory for contrastive representation learning will be presented at ICML on July 7. This work analyze the global contrastive objective, rather than the mini-batch contrastive objectives used in prior theoretical analyses, which are inherently flawed.
1/2: my book "Compositional Optimization for Advanced Machine Learning" is heading to Springer for production! 📖
SGD and Adam are not the whole story. Many modern ML objectives are intrinsically compositional.
👉 opt4ml.org 🚀
2/2 including softmax, contrastive learning, AUC maximization, and ranking. This book covers it all — from classical stochastic optimization to the unified compositional framework needed for modern AI: theory, algorithms, and real-world applications.
PDF: opt4ml.org/assets/compositio…
1/4 🧠⚡ Can reasoning models think faster without thinking worse?
Recent systems like Meta Muse Spark use length penalties in training to reduce unnecessary output tokens.
A simple approach is to combine penalty-based reward with correctness-based reward in GRPO.
3/4 🚀 In our recent ICLR paper, we propose DRPO: Decoupled Reward Policy Optimization.
💡 Key idea: decouple length optimization for correct and incorrect rollouts, so the model learns to be concise without punishing valid reasoning.
4/4 📊 The results are remarkable: DRPO surpasses six strong GRPO-based baselines, achieving much shorter reasoning traces while maintaining or even improving performance.
📄 Paper: arxiv.org/abs/2510.04474
💻 Code: github.com/Optimization-AI/D…
Tianbao (TB) Yang retweeted
Thank you so much for the shoutout!! @GalantiTomer
If you’re excited to dig into why self-supervised contrastive learning works so well, come swing by our poster session!
📍 Exhibit Hall C/D/E — Poster #2607
🗓️ Fri, Dec 5 • 4:30–7:30 PM PST
Sadly, I won't be at @NeurIPSConf this year. But my brilliant Ph.D. student @luthra_achal will be there. If you are curious what is the relationship between self-supervised and supervised learning, check out our poster!
1/2 Given the recent leakage of reviewer identities, some authors have reported that their papers were reviewed by people from the same institution or even the same research group. This raises serious concerns: how is such a conflict of interest even possible?