Lang Huang

Assistant Professor & PI at NII (@jouhouken), RS@Sakana AI | Ph.D. at UTokyo | Ex-Intern@Adobe, and SenseTime | Working on self-improving machine learning.

[Alaya Lab @alayastd]

This month, Alaya Lab is releasing a series of research projects and open-source initiatives.
Today, we're excited to introduce Alaya World - an open-source interactive video world model.
Project Page:https://t.co/cyVypvewZh
Github: https://t.co/8BKKpzHPq5
✨ Highlights:
- 720p 24 FPS streaming generation
- Navigation and prompt-driven interactions (e.g., spell casting and summoning)
- Stable long-horizon generation (>1 minute)
- State-of-the-art performance
- Inference code is available today. Training code and datasets are coming soon.

[Lang Huang @LayneHuangLH]

Thanks for sharing our paper. Please check out our project page for the interactive demo and more! https://t.co/jux3xJS9AL

[Valeriy M., PhD, MBA, CQF @predict_addict]

The most underrated skill in forecasting?

Not prompt engineering.
Not model stacking.
Not chasing the newest architecture.

It’s tuning ridge regression apparently.

A few years ago, a paper showed that simple linear neural networks could outperform Transformers on time series forecasting.

Since then, the “Transformers will solve forecasting” narrative has never really recovered.

Fast-forward to 2026, and the lesson feels even clearer:

For many time series problems, the model you need is not bigger.
It is not deeper.
It is not more attention.
It is a well-tuned linear baseline.

Ridge regression remains boring, fast, interpretable, and annoyingly hard to beat.

And in forecasting, “boring and hard to beat” is often exactly what wins.

The future of time series may not be Transformer-shaped.

It may be regularized.

#timeseries #forecasting

[Valeriy M., PhD, MBA, CQF @predict_addict]

The most underrated skill in forecasting?

Not prompt engineering.
Not model stacking.
Not chasing the newest architecture.

It’s tuning ridge regression apparently.

A few years ago, a paper showed that simple linear neural networks could outperform Transformers on time series forecasting.

Since then, the “Transformers will solve forecasting” narrative has never really recovered.

Fast-forward to 2026, and the lesson feels even clearer:

For many time series problems, the model you need is not bigger.
It is not deeper.
It is not more attention.
It is a well-tuned linear baseline.

Ridge regression remains boring, fast, interpretable, and annoyingly hard to beat.

And in forecasting, “boring and hard to beat” is often exactly what wins.

The future of time series may not be Transformer-shaped.

It may be regularized.

#timeseries #forecasting

[Lang Huang @LayneHuangLH]

7/7 Paper, code, and more examples:
arXiv: https://t.co/qMJQ22cSb1
Code: https://t.co/KG6J9JM0Va
Project: https://t.co/tBVy4L9uPs

This is joint work with Yonghao Yu, Ruiyi Li, Zerun Wang, and @toshi_yamasaki.

Enjoy!

[Lang Huang @LayneHuangLH]

1/n Video-Mirai: Autoregressive Video Diffusion Models Need Foresight.

Causal video generators decode one segment from only the past. With no representation of where a scene is heading, subjects and backgrounds drift.

We give them foresight, at training time only. 🧵

[Lang Huang @LayneHuangLH]

6/n A probe makes the mechanism visible. A frozen readout reconstructs future frames from the hidden state alone, and it matches the average of multiple stochastic rollouts. The state appears to encode a distribution of plausible futures, not one trajectory.

[Lang Huang @LayneHuangLH]

I will be presenting Mirai at #CVPR2026 today.
📍: ExHall F #176
🕰️: 11:45 AM – 13:45 PM MDT

Come to our poster for a chat!

[Lang Huang @LayneHuangLH]

1/n Autoregressive Visual Generation Needs Foresight.

Excited to share our work on autoregressive image generation, Mirai, accepted at #CVPR2026.

[Lang Huang @LayneHuangLH]

@goqh_so Thanks!

[Lang Huang @LayneHuangLH]

1/n Autoregressive Visual Generation Needs Foresight.

Excited to share our work on autoregressive image generation, Mirai, accepted at #CVPR2026.

[Lang Huang @LayneHuangLH]

5/5 Huge thanks to all the co-authors, Yonghao Yu, Zerun Wang, Runyi Li, and @toshi_yamasaki .

This kicks off our line of work on autoregressive modeling and visual generation. More coming soon.

#CVPR2026 #GenerativeAI #ComputerVision

[Lang Huang @LayneHuangLH]

4/n Project page: https://t.co/9KUnJdYOkQ
Paper: https://t.co/TPM0zLyeg0
Code: https://t.co/cRzx9jIGHr

[Lang Huang @LayneHuangLH]

So cool

[Alaya Lab @alayastd]

𝗜𝗻𝘁𝗿𝗼𝗱𝘂𝗰𝗶𝗻𝗴 𝗚𝗲𝗻𝗲𝗿𝗮𝘁𝗶𝘃𝗲 𝗪𝗼𝗿𝗹𝗱 𝗥𝗲𝗻𝗱𝗲𝗿𝗲𝗿
A new toolkit, dataset, and baseline for world-scale rendering:
- collect G-buffers from AAA games
- scale data for rendering complex world scenes
- improve rendering performance
- enable game effect editing

[Ryan Yuan @RainbowYuhui]

We’re excited to launch #MagicLayers globally at Canva 🚀🚀🚀

This is currently the most advanced image-to-layer decomposition model in the world-purpose-built for design-native content.

⚡ 20–200× faster than Qwen-Image-Layer
🎨 Super strong on designs

https://t.co/vQ7cCPnGGZ