Zhenglin Wan

CS Ph.D @NUSingapore | Prev: @CUHKofficial | National Chess athlete | Building multi-modal intelligence • Share insights

[Zhenglin 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: https://t.co/mBjBE4b3ps
💻 GitHub：https://t.co/1mtCdu7cxT
🤗 Hugging Face Papers：https://t.co/OLrLjZ95KV

After nearly 800 wandb runs, we found some surprising answers. 🧵👇

[Yuwei Niu @purshow04]

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.

[Zhenglin Wan @Vanzllll]

Thanks for the shout-out to DiffusionART! Glad to have W&B along the way 😄 Excited to try the LEET updates!

[Weights & Biases @wandb]

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

[Zhenglin Wan @Vanzllll]

@wandb Thanks for the shout-out to DiffusionART! Glad to have W&B along the way 😄 Excited to try the LEET updates!

[Zhenglin Wan @Vanzllll]

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.

[Zhenglin 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: https://t.co/mBjBE4b3ps
💻 GitHub：https://t.co/1mtCdu7cxT
🤗 Hugging Face Papers：https://t.co/OLrLjZ95KV

After nearly 800 wandb runs, we found some surprising answers. 🧵👇

[Xuhong Huang @cai_xu77719]

https://t.co/A1dlSj77cN
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.

[Zhenglin Wan @Vanzllll]

@xzhang_billy 是的，diffusion是薄肌模型，某种程度上也弥补了。

[Zhenglin 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: https://t.co/mBjBE4b3ps
💻 GitHub：https://t.co/1mtCdu7cxT
🤗 Hugging Face Papers：https://t.co/OLrLjZ95KV

After nearly 800 wandb runs, we found some surprising answers. 🧵👇

[Zhenglin Wan @Vanzllll]

@qingke_ai Thanks for reposting！

[Zhenglin Wan @Vanzllll]

Our analysis suggests it's not primarily about reward scale.
Interestingly, both branches provide aligned reward guidance, but their hidden forward–reverse consistency forces point in opposite directions.
Joint training partially cancels these forces, while single-branch training leaves a strong residual that can cause gradient spikes and destabilize training.
So the key issue is the uncanceled consistency force, rather than simply insufficient reward signal.

[Zhenglin Wan @Vanzllll]

💡 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: https://t.co/mBjBE4bBf0
💻 Code: https://t.co/1mtCdu7Knr
🤗 HF Papers: https://t.co/OLrLjZ9DAt
⭐ Stars, upvotes, and feedback are greatly appreciated!

[Zhenglin Wan @Vanzllll]

🔍 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.

[Bojie Li @bojie_li]

我要改变一下对 Jev 类模型的观点。刚发布的时候我觉得它不过是个分类模型，现有 LLM decode 一个 token 就行了。现在意识到它技术上不难，关键是 System 1 的思想：世界是事件驱动的，每次只做一点简单决策。

首先，之前 LLM 是用在 agent react loop 里面的，外围的 harness 和业务逻辑里尽量不用 LLM，否则成本和延迟都��不了。Jev 和今天刚发布的 Pine Computer 的逻辑是，只要模型够快、够便宜，那么很多原本只能用 if-else 凑合的业务逻辑都可以上模型了。

更深一步想，人的实时决策其实更像 System One 模型，而不像 autoregressive LLM。

- 在语音对话中，人每几百毫秒都要做出打断、附和、停顿等交互判断，这是 AI 说话像人的关键。在 GPT-Live 1 和 TML interaction model 这些 full duplex model 中，每个几百毫秒的时间片它只输出几个 token。

- 在 computer use 中，人每几百毫秒都要决定下一步点哪里，UI 一改变，本来准备点的位置也不点了。Pine Computer 就是把 computer use 从 turn based 变成 event driven 的了，输出也尽可能从坐标简化成了 label。

- 在 robotics 中，人每几百毫秒都在规划下一步的动作，或者停止当前动作。

在交互中，观察反馈循环的频率很可能比智商重要。Full-duplex models、Jev 和 Pine Computer 都在指向一个新的方向，SOTA background model 负责深度思考、解决难题，而 System One Models 负责实时语音交互、computer use、业务逻辑判断，甚至未来的具身操作。

[Zhenglin Wan @Vanzllll]

If generation cannot help understanding, why we need Chain of Thought（CoT)？LLM should just output the answer instead of a long reasoning chain.

[Zhenglin Wan @Vanzllll]

I do think the 60000 Iclr submission shows that the era of researcher is coming back, as research itself can be disentangled from paper outputs and this evaluation criteria, and the researcher can pursue the groundtruth.

[Phillip Isola @phillip_isola]

Students sometimes ask me if it still makes sense, in this accelerating age, to pursue a PhD in AI.

Perhaps counterintuitively, I think it's a great time to do so.

I wrote up some thoughts on this here: https://t.co/RRqGR0qi1R

[Zhenglin Wan @Vanzllll]

@phillip_isola I think the 60000 Iclr submission shows that the era of researcher is coming back, as research itself can be disentangled from paper outputs, and the researcher can pursue the groundtruth.

[Zhenglin Wan @Vanzllll]

@joielovecrypto Hi happy ro to connect ！

[Zhenglin Wan @Vanzllll]

@pixel2ai Great

[Zhenglin Wan @Vanzllll]

Astra has already solved GUI (and robot use is 3D version of GUI)

[Xuhong Huang @cai_xu77719]

I’d recommend everyone working on multimodality to read this Roboflow post on GPT-6 Astra:
https://t.co/eyyLp3poWi

After working on multimodality for half a year, I increasingly feel that some of the most important problems are still the old ones:

OCR. Grounding. Segmentation.

They rarely make the fanciest demos, but they are the foundation of everything downstream.

2D grounding is fundamental to Computer-Use Agents; 3D grounding will be fundamental to Robot-Use Agents.

These capabilities need both targeted training and end-to-end optimization inside agents. Astra is one of the clearest examples I’ve seen of the two coming together.

And that is exactly why Astra can pull off so many fancy Blender and robot-control demos: the flashy behavior is built on very unflashy fundamentals.

Multimodality is not the destination. In the agent era, perception matters because it enables action.

Everything ultimately comes down to taking action.

[Zhenglin Wan @Vanzllll]

@cai_xu77719 Agree. Grounding is one of the most fundamental problem