Zhenglin (Carlos) Wan (@Vanzllll) header
Zhenglin (Carlos) Wan (@Vanzllll) profile picture

Ph.D @ NUS HPC-AI Lab | https://xitter.cf/t.co/sPSzz2qNlN @ CUHK Multimodal Understanding & Generation National Chess athlete | CMOer

Singapore
Joined December 2025
Photo by Zhenglin (Carlos) Wan (@Vanzllll)Photo by Zhenglin (Carlos) Wan (@Vanzllll)Photo by Zhenglin (Carlos) Wan (@Vanzllll)Photo by Zhenglin (Carlos) Wan (@Vanzllll)Photo by Zhenglin (Carlos) Wan (@Vanzllll)Photo by Zhenglin (Carlos) Wan (@Vanzllll)Photo by Zhenglin (Carlos) Wan (@Vanzllll)
🚀 Introducing DiffusionART: efficient & stable & Scalable diffusion RL. Understanding forward-process RL first: 1. ❓Positive branch and negative branch provides the same ideal gradient direction, but why do reward collapses without any of them? 2. ❓Why does EMA off-policy strategy stablize training? ⚡ On HunyuanVideo-13B T2V: DiffusionART is 12.02× / 11.72× more efficient than DiffusionNFT and Flow-GRPO-Fast. 📄 Arxiv: arxiv.org/abs/2610.12572 💻 GitHub:github.com/vanzll/DiffusionA… 🤗 Hugging Face Papers:huggingface.co/papers/2610.1… After nearly 800 wandb runs, we found some surprising answers. 🧵👇
4
6
4
69
10,362
Zhenglin (Carlos) Wan retweeted
Replying to @JensenHuang
@JensenHuang, mission is the boss. Your mission in re-industrializing USA through Physical AI is not yet done. We demand your return to get the job done.
Over the past week, I’ve been overwhelmed by the love and support of friends, industry colleagues, and people I've never met. Thank you to everyone who has shown kindness to my family and me. It means the world to me. I came to the United States 53 years ago, and becoming an American citizen is one of the greatest gifts and proudest moments of my life. NVIDIA has been my life’s work. Building it into the company it is today alongside my colleagues has been a privilege and an honor. Nowhere but in America could a company like NVIDIA be possible. I love our country, and I love NVIDIA. I did not live up to the standards expected of everyone at NVIDIA, and I have made serious mistakes in my personal life. I created problems for the Board that I deeply regret. I own my mistakes completely, and I am committed to cooperating with the Board’s independent investigation. Since my departure, I’ve seen an outpouring of calls from NVIDIA’s employees, investors, business partners, and customers for me to return as CEO. Many continue to reach out to me directly as the company navigates complex technology, market, competitive, and geopolitical dynamics. NVIDIA is a once-in-a-generation company. If the Board of Directors decides that my return is in the best interest of our shareholders, our employees, the company, and America, I would be honored to answer these calls and return as CEO. Thank you again to everyone who has shared their good wishes. I am profoundly humbled by your support. To my family: I love you, and I will spend the rest of my life being worthy of yours. Jensen
21
56
2
496
17,320
Zhenglin (Carlos) Wan retweeted
Over the past week, I’ve been overwhelmed by the love and support of friends, industry colleagues, and people I've never met. Thank you to everyone who has shown kindness to my family and me. It means the world to me. I came to the United States 53 years ago, and becoming an American citizen is one of the greatest gifts and proudest moments of my life. NVIDIA has been my life’s work. Building it into the company it is today alongside my colleagues has been a privilege and an honor. Nowhere but in America could a company like NVIDIA be possible. I love our country, and I love NVIDIA. I did not live up to the standards expected of everyone at NVIDIA, and I have made serious mistakes in my personal life. I created problems for the Board that I deeply regret. I own my mistakes completely, and I am committed to cooperating with the Board’s independent investigation. Since my departure, I’ve seen an outpouring of calls from NVIDIA’s employees, investors, business partners, and customers for me to return as CEO. Many continue to reach out to me directly as the company navigates complex technology, market, competitive, and geopolitical dynamics. NVIDIA is a once-in-a-generation company. If the Board of Directors decides that my return is in the best interest of our shareholders, our employees, the company, and America, I would be honored to answer these calls and return as CEO. Thank you again to everyone who has shared their good wishes. I am profoundly humbled by your support. To my family: I love you, and I will spend the rest of my life being worthy of yours. Jensen
60,510
121,582
24,445
1,171,944
57,821,717
In Chinese, it's called "大道至简,无欲则刚"
《其实我们不需要那么多理论》 在这样一个基模吞并一切的时代,我非常理解大家作为 新时代PhD 的焦虑与无力感。看着推上每天都在诞生新的生存法则,大 Bench、大实习、大粉、大日结、大斜率、HR 理论…… 但冷静下来想想,我们真的需要这么多“理论”吗? 1.有时我们真不需要那么多理论,有什么想做的 idea,直接干就好。 确保你跟上时代、不要做太牢的 topic 就行,真没必要一股脑地跟风做 bench、做 agentic 具身。因为好的 work 是需要 commit 大量时间的,等一个东西已经被大家总结成“理论”火起来了,你才开始做,说明你已经做晚了。跟随着自己感兴趣的topic一直做下去,才能让你对specific area有某种先于他人的"priors",即good taste 2. 基模越是吞噬一切,Taste反而越贵。 如果你做的事情能被一套简单的“大 XX 理论”概括,那明天 RSI 和基模就能比你做得更快更好。当所有人都在日结的时候,愿意做一件按年结算的事,本身就是稀缺。“能一直被高价挖的 是身上有不按天结算的东西的人” 3. 我们之所以热衷于给一切贴上“理论”的标签,是因为读博最大的痛苦就是直面巨大的不确定性。但可惜的是人生没有这样的rl飞轮。 关掉所有的理论,找个自己真正好奇、哪怕稍微有点笨的问题,扎进去做点自己回过头看不会觉得羞愧的东西 --- Rethinking 大xx理论
1
173
Zhenglin (Carlos) Wan retweeted
其实心态放平点就好,转LLM不一定是立马放弃去投个基模组,这样也很难有能做有意义事情的team去接收你。这其实是个需要运营的过程(至少有认识的人帮忙推荐,至少需要了解大家在关心什么) 把各个环节(arch design、optimizer、parallelism)的很多技术当成有趣的东西来学习,多involve一些benchmark(或者投奔柴哥)。其实也就挺好的,给自己找点事情干。 选择是一种处境,如果真的纠结短暂利益和长远发展。那就按照着折中方案亦步亦趋,等待情况变化再走向下一步。 人生的容错率很高,如果你愿意拼一阵子那你就持续努力地向最核心的内容靠齐;如果想休息了就苟着,多学习多思考。
In fact, the rise of LLMs began as early as 2023 (even 2022). Coming from autonomous driving, I've seen researchers spend years trying to replicate LLMs' success in vision, while others simply turned to LLMs themselves. I respect those who still hold a genuine belief in vision. But I suspect many just don't want to leave their comfort zone.
6
6
2
119
10,405
Zhenglin (Carlos) Wan retweeted
The U.S. wouldn’t be leading in AI today without @JensenHuang. His push into physical AI aims to reindustrialize America and keep it ahead for decades. Ousting him would undermine U.S. technological leadership and work against the national interest.
27
87
7
755
26,137
我有不一样的观点, 计算机图形学(cg)的上限就是code for video gen的上限。fable现在本质上还是在调用现有的rule-based渲染器
#Claude Fable 5.2(即将)扑灭的领域有 1. Video Generation: xitter.cf/alexmichael/status/211… 2. 3D Generation: xitter.cf/petergostev/status/211… 3. Game Design: xitter.cf/felixrieseberg/status/… 4. Robotic Manipulation: xitter.cf/claudeai/status/211986… xitter.cf/karolina_dubiel/status… Visual Coding 从被提出到被 kill 也不到半年
7
3
1
42
3,261
昨天我还在和@FanqingMengAI 讨论,如何抵抗AI时代的欲望,焦虑。 金钱, 名利, 都是身外之物, 我很多时候享受现在拥有的平庸和平淡,至少我还拥有最宝贵的东西:健康, 自由, 亲情, 友情。我一直臣服于命运, 并对现在拥有的一切感恩。 Best wishes to Weiyang 🙏🙏
Announcement: I have indeed been diagnosed with the disease that everyone fears, but it was caught at a very early stage. I have already scheduled my surgery, and I thank everyone for your concern. Honestly, I no longer care about so-called money or prestige. I just want to do what I can within my academic field and work on AI things that bring me joy (the doctor says this might also help with my recovery). I have let go of all past interpersonal entanglements, as well as others' prejudices and discriminations against me. In the face of life and death, these things are so insignificant. It is this illness that has made me rethink the meaning of life and death: one should not live for the approval of others, but persist for the things we love and the people who love us. I will continue to share because it is a meaningful thing to do.
2
25
1,243
Introducing our latest research The Trinity: a capability-driven multimodal RSI, from self-generated worlds to real robots. We started with two questions: 1. Are tasks the right unit of evolution for RSI? 2. How can world generation and capability evolution reinforce each other in a sustained feedback loop? ⚡ With model weights frozen: average gains of +34 points across 28 training benchmarks and +26 across 18 unseen benchmarks. 🤖 Real2Sim2Real: success on both Unitree G1 and AlohaMini2 tasks improves from 0% → 60%. 🌐 Project: trinity-rsi.github.io 📄 arXiv: arxiv.org/abs/2610.36077 💻 GitHub: github.com/InfRec-Cardinal/T… 🤗 HF Papers: huggingface.co/papers/2610.3… Our findings point to shared capabilities as a foundation for transferable self-improvement.
1
1
11
1,590
Finally, the grand finale of InfRec Launch Week: Route Anything. A 4B router orchestrates a system that beats fixed Opus 5.5 on our 36-input Terminal-Bench test: 77.78 vs 69.44. The best single model isn't the ceiling. arxiv.org/abs/2611.05275
2
5
470
1/5 · RELEASE ⚡ DiffusionART-Flash × MiniMax-H3 We've open-sourced ART-Flash and our MiniMax-H3 RL post-training recipe. WHAT'S NEW • Train on just 1–2 high-noise timesteps • Comparable reward gains and stability in our experiments • For a 50-step schedule: 25× fewer sample-timestep pairs when training on 2 steps Rollouts still follow the complete sampling trajectory. THE H3 RECIPE • 480P · ~5 seconds · 10 rollout steps • Video: train on 2 of the first 4 steps • Audio: train on 2 of the first 3 steps • Six open-source rewards, normalized within prompt groups and routed by modality 📊 Recipe + training curves attached 🧵 Next: three technical answers + seven engineering lessons 🔗 github.com/vanzll/DiffusionA… Thanks to Synced (机器之心) for the coverage, and to everyone sharing questions and ideas. If this helps your research, a GitHub star is appreciated!
1
5
1
27
2,640
4/5 · GLOBAL COVARIANCE Why use global covariance instead of local posterior covariance? THE THEORETICAL ROUTE Continuing DiffusionNFT's derivation would require a covariance under the posterior over x₀ given xₜ induced by the forward noising dynamics. Computing or estimating it adds substantial complexity and makes scaling harder. THE ART CHOICE ART is inspired by NFT, but does not continue its derivation. We use the population-mean intuition in 2/5 and systematically remove the forward–reverse consistency force. Global covariance keeps the method simple and scalable.
1
1
60
5/5 · ENGINEERING NOTES ① Prompt batch size Aim for at least 16–32 prompts per update. Too few prompts can cause high gradient variance and stalled rewards at otherwise reasonable learning rates. Save memory with sequential rollouts, smaller microbatches and gradient accumulation; preserve the effective prompt batch. ② Multi-reward normalization Normalize each reward within its prompt group before combining rewards with different scales. ③ MiniMax-H3 resolution Very low resolutions put rollouts far out of distribution. For simpler prompts, fewer rollout steps (e.g. 10) can save time and memory. Our recipe uses 480P. Below that, rollout quality and training degrade substantially. Use the recommended 768P when resources permit. ④ Modality-specific routing For an omni model with modality-specific reward models, compute covariance and masks separately for each modality, then average the losses. ⑤ Cache rollout outputs Avoid an extra forward pass per state during updates. Caching and training only the conditional CFG branch can save another forward pass and substantial GPU time. ⑥ Reward quality and hacking In our experience, open-source video rewards still fall short of human judgment and are prone to hacking across algorithms. For algorithm research, I treat the chosen reward as the optimization objective. Reward curves alone do not establish human-perceived quality. In our experience, closed-source video rewards tend to be stronger. ⑦ SDE rollouts ART also supports SDE rollouts. Group-based methods suffer when diversity is too low. SDE sampling can increase diversity, with ART guidance applied to the SDE mean. Full recipe → github.com/vanzll/DiffusionA…
1
81
Zhenglin (Carlos) Wan retweeted
I'm increasingly convinced that MAI-Thinking-1 and Qwen3.8-Flash-Next are two must-read technical reports for anyone who isn't on a frontier pretraining team but wants to learn the practical know-how behind training frontier models.
11
57
2
1,073
51,322
Thanks for the shout-out to DiffusionART! Glad to have W&B along the way 😄 Excited to try the LEET updates!
wandb 0.31.0 is out now, and import wandb and the CLI start about twice as fast. That's for builders like @Vanzllll, whose DiffusionART team "found some surprising answers" after nearly 800 wandb runs. Now find yours. It also grows W&B LEET, the terminal UI for exploring W&B runs. wandb leet inspect --summary reads a run's local .wandb file, and --json prints that as JSON for a coding agent. pip install -U wandb
2
1
1
12
1,332
DiffusionART has received a lot of attention since public, and I am really happy to hear about that the some frontier VideoGen companies of China have been trying to put it to use for their post-training pipeline and get a good result. I'd like to provide more insight about the meaning of "aggregation-first" since some researchers reached out to me for discussing this confusing term: NFT-style guidance to each state contains an harmful consistency force and the dual branch cancel them only partially via cross-sample gradient aggregation (a.k.a loss = mean(per_sample_loss)). ART-style training uses the Covariance of endpoint and reward to aggregate a clean reward-improving direction before applying this guidance to each state, systematically cancelling the harmful force. Interestingly, we can regard DiffusionART as the combination of Forward-Process RL and Reverse-Process RL. Meanwhile, the term "guidance" is inspired by CFG (classifier-free guidance), and shares some basic principle. So another way for interpretation is ART is "internalizing" the "reward-improving guidance" to model weight via MSE loss.
🚀 Introducing DiffusionART: efficient & stable & Scalable diffusion RL. Understanding forward-process RL first: 1. ❓Positive branch and negative branch provides the same ideal gradient direction, but why do reward collapses without any of them? 2. ❓Why does EMA off-policy strategy stablize training? ⚡ On HunyuanVideo-13B T2V: DiffusionART is 12.02× / 11.72× more efficient than DiffusionNFT and Flow-GRPO-Fast. 📄 Arxiv: arxiv.org/abs/2610.12572 💻 GitHub:github.com/vanzll/DiffusionA… 🤗 Hugging Face Papers:huggingface.co/papers/2610.1… After nearly 800 wandb runs, we found some surprising answers. 🧵👇
1
1
14
1,633
Zhenglin (Carlos) Wan retweeted
arxiv.org/abs/2605.27284 Thats what we do in this work, FineVLA. The upper bound of VLA might be the image/video gen, which is very good when the prompt is quite well (usually polished with a llm refiner) And until now, the only one intelligence I see like human is the CoT lol.
Really agree with what @JiachenAI said at Alignment 2026 today. One possible reason VLAs don't generalize well is how the data is presented. We train VLAs to map short text instructions to long, complex trajectories. But the instructions often don't contain enough information to explain the trajectories. As a result, the model tends to overfit to other details in the environment rather than the instructions. Maybe we should try much richer text instructions that describe how to actually perform the task. This reminds me of image and video generation. The core generation models still rely on increasingly detailed prompts. The job of turning a short user request into a detailed prompt is now handled by an LLM or agent. Perhaps VLAs could benefit from a similar approach. Sometimes I feel language really is one of humanity's greatest inventions.
1
5
332
🔍 Two informative insights: 1️⃣ Hidden Consistency Force NFT-style forward-process objectives contains an additional term: a forward–reverse consistency force beyond reward guidance. Why we need both positive and negative branches? Dual-branch oppose this force, and heuristically cancel this unwanted force. But can only cancel it by cross-sample gradient aggregation—not state by state—leaving destabilizing residuals and gradient spikes. 2️⃣ Why EMA Really Works EMA stabilizes forward RL through old-policy restoration in the loss, NOT off-policy rollouts themselves. Surprisingly, this restoration force can dominate parameter gradients despite being much weaker in output space! The fundamental difference between on-policy and off-policy training in continuous-space diffusion RL is restoration. Off-policy training can provide restoration.
1
500
💡 Building on these insights, we introduce DiffusionART: Aggregate reward guidance first, restore as the policy moves. ① Aggregation-First Guidance: Aggregate reward signals into a shared reward-improving direction via Cov(x₀, r), then guide actual reverse-ODE rollout states using frozen MSE targets—eliminating the hidden consistency force by construction. ② Adaptive Restoration: A trust-region-masked second update introduces drift-adaptive restoration, without EMA tuning! ✅ No ODE-to-SDE conversion ✅ No differentiable reward model required ✅ Supports black-box & pairwise rewards ✅ Robust to smaller prompt group sizes ✅ Stable and efficient across image & video generation, without per-task tuning 📈 On HunyuanVideo-13B, DiffusionART achieves +28.8% / +53.0% higher mean training rewards vs DiffusionNFT / Flow-GRPO-Fast. 🔓 Everything is open-sourced! 📄 Paper: arxiv.org/abs/2610.12572 💻 Code: github.com/vanzll/DiffusionA… 🤗 HF Papers: huggingface.co/papers/2610.1… ⭐ Stars, upvotes, and feedback are greatly appreciated!
323