Chieh-Hsin (Jesse) Lai

This account is verified. Learn more

@JCJesseLai

Tokyo-to, Japan chiehhsinjesselai.github.io Joined September 2022

439 Following 2.8K Followers 515 Posts

Latest posts 5

Pinned post

Chieh-Hsin (Jesse) Lai @JCJesseLai Oct 29, 2025

Tired to go back to the original papers again and again? Our monograph: a systematic and fundamental recipe you can rely on!

📘 We’re excited to release 《The Principles of Diffusion Models》— with @DrYangSong, @gimdong58085414, @mittu1204, and @StefanoErmon.

It traces the core ideas that shaped diffusion modeling and explains how today’s models work, why they work, and where they’re heading.

🧵You’ll find the link and a few highlights in the thread.
We’d love to hear your thoughts and join some discussions!

⚡ Stay tuned for our markdown version, where you can drop your comments!

Chieh-Hsin (Jesse) Lai @JCJesseLai Oct 22

🎉 Excited to bring our book The Principles of Diffusion Models to @NeurIPSConf 2026 as a tutorial—this time, all about diffusion LLMs! 🚀

🔹 How do ideas from continuous diffusion extend to language generation?

🔹 What are the latest advances and open frontiers in diffusion LLMs?

From foundational principles to cutting-edge research. Stay tuned!

📆 Mon, Dec 7, 9 AM AEDT
📍 TikTok Ent Centre (Sydney, Australia)

w/ organizers @ssahoo_ @mittu1204 @StefanoErmon

Subham Sahoo Verified account @ssahoo_ Oct 22

🎊 Excited to give a tutorial on diffusion LLMs at @NeurIPSConf 2026 with @StefanoErmon @JCJesseLai @mittu1204

📆 Mon, Dec 7, 9 AM AEDT
📍 TikTok Ent Centre (Sydney, Australia)

Topics covered:
> A beginner friendly introduction to diffusion-LLMs
> Designing frontier diffusion-LLMs
> Using Continuous diffusion for LLMs

View quoted post on X

Chieh-Hsin (Jesse) Lai @JCJesseLai Oct 1

⏰ How much does a Uniform Discrete Diffusion Model actually need time?

Our answer: often, surprisingly little.

The interesting tension is this:

🤔 The optimal UDM predictor is theoretically time-dependent.

But on finite language data, the noisy sequence itself can already reveal much of what the model needs to know about which tokens are reliable.

And empirically, the result is interesting:

🚀 Removing time conditioning does not hurt performance, and can even improve it.

So while time matters in the idealized theory, practical UDMs may rely on it far less than expected.

See more in the quoted post and our paper 👇

w/ @ChunsanHong, @takiko_san, @kurowassann621, @mittu1204

Chunsan Hong Verified account @ChunsanHong Oct 1

⏰ Does Uniform Discrete Diffusion Need Time?

Surprisingly, often NO.
📐 In theory, the optimal predictor depends on time.
📉 But with finite data, this dependence can largely disappear.

And in practice, time-agnostic UDMs can work just as well, or even better!!

🧵 w/@JCJesseLai, @kurowassann621, @takiko_san, @mittu1204
View quoted post on X

Chieh-Hsin (Jesse) Lai @JCJesseLai Oct 1

Happy to share a new work done together with my intern @bskim98.

🌏 World models don’t just need more memory. They need the right memory to predict what happens next.

🔮 We introduce FAR (Future-Aware Recall), a framework that learns which useful cues to trust among multiple available cues for episodic recall in world models, based on a simple principle.

See more in the quoted post!

w/ @nguyencongbacbk @_amirbar @mittu1204

Beomsu Kim Verified account @bskim98 Oct 1

(1/9)🧠World models don’t just need more memory. They need the right memory to predict what happens next.

We introduce FAR (Future-Aware Recall), a principled framework that learns episodic recall for world models:

🏆Beats hand-designed recall—even with the same cues
🎛️Automatically learns which recall cues to trust for each query, among all available cues
🔄Recalls the right state, even as the environment changes

Huge thanks to my collaborators @JCJesseLai, @nguyencongbacbk, @_amirbar, Jong Chul Ye, and @mittu1204 for making this work possible!

View quoted post on X

Chieh-Hsin (Jesse) Lai @JCJesseLai Sep 8

If the claimed finite-time blow-up result for 3D incompressible Navier–Stokes with smooth, compactly supported forcing survives independent verification, it would establish Clay’s C/D formulation, while leaving unforced Navier–Stokes open. Either that, or a future resolution of the unforced problem, would leave me genuinely shocked and conflicted.

One side of me is deeply excited by the pace of AI progress. I did not imagine AI could reach this level of mathematical discovery and proof construction so soon.

The other side, as someone trained in mathematics, feels for today’s math students. A problem they spend years pondering could suddenly be cracked once an AI lab directs enough compute toward it. Otherwise, it might remain open simply because nobody with those resources considers it important enough to solve.

There is also the coming flood of AI-generated mathematics. Genuine breakthroughs may be mixed with huge amounts of polished but subtly wrong work. Producing plausible proofs could become cheap, while understanding and verifying them remains expensive. That could become a serious headache for the entire mathematical community.

OpenAI Verified account @OpenAI Sep 8

We’re sharing a solution to the Navier-Stokes Millennium Prize Problem, one of the deepest problems at the frontier of mathematics.

The proof was produced by a group of agents, using an OpenAI next-generation model significantly more capable than GPT-6 Astra.

The problem concerns whether the description of smooth three-dimensional fluid motion modeled by the Navier-Stokes equations can break down. It has remained unresolved for roughly 90 years.

View quoted post on X