Zihan Ding

PhD @Princeton University | Research @Bytedance Seed Multimodal | Previous research @Meta GenAI / Adobe Research / Meta FAIR / Tencent Robotics X / Borealis AI

[Zihan Ding @Hanry65960814]

Finally understood why some industrial people don’t appreciate good research taste anymore. Bypassing the hard problems to meet the delivery due is survival rule under both internal and external pressure sometimes..

[Zihan Ding @Hanry65960814]

The current paper review system is a failure, and it will be more obvious as AI-centric research develops. The quality degrades from journals to conferences, and a huge degradation since AI exploded. We should customize an AI review system (with code verification) for peer review (efficient and higher quality) even to review the AI generated papers. Very soon both sides will be fully automated.

From industry perspective (at least from my view), 99% work are wrong (mostly wrong at beginning from its direction) and 90% are meaningless (excluding for educational purpose), but we count on the 1% to progress.

This also implies a wrong academic credit/scholar system. Why a paper in related work but not so relevant has same citation credit as those few critical baselines/ideas in one paper?

[Xudong Han @Xudong07452910]

这篇普林斯顿博士论文，几乎把强化学习过去十年的两条主线串在了一起：

从博弈算法，一路走到世界模型。

作者 Zihan Ding，导师是 Chi Jin。整篇论文超过 300 页，汇集了博士期间八项研究工作，其中多项成果发表于 NeurIPS、ICML、ICLR 和 ICCV。

前半部分研究多智能体博弈：如何学习接近纳什均衡、难以被对手利用的策略，并把这些方法扩展到格斗游戏、多人博弈和真实网络负载均衡。

后半部分转向 Foundation Model 时代：扩散世界模型、生成模型策略、少步视频生成、可交互视频世界模型，以及支持长时预测的记忆机制。

这篇论文最有价值的地方是它记录了强化学习研究重心的一次迁移。

过去，RL 更关注如何在给定环境中学会行动。

现在，Agent 还需要借助生成模型建立对环境的理解，预测不同动作会带来什么未来，再据此规划和决策。

这里收录的也不只是零散的想法，而是一条经过多年研究和多篇顶会论文逐步铺开的路线。

很多论文告诉你某个方法如何提升一个指标。

这篇博士论文更像一���地图，让人看清强化学习如何从策略优化，走向对复杂世界的建模。

想系统理解 RL、世界模型和生成式 AI 之间的关系，这篇很值得完整读一遍。

📎 arxiv: https://t.co/M7rIJ6SKwO

[Zihan Ding @Hanry65960814]

More towards unified multimodal

Representation Forcing: https://t.co/ZXduuGuq4d

Context Unrolling: https://t.co/8xr3D6Ei4u

[Zihan Ding @Hanry65960814]

Why no such joint design on robotic arms

[1517 Fund @1517fund]

We live in an age of miracles.

3D printing and computational geometry make previously unimaginable geometries possible.

From: Kazuki Abe, Riichiro Tadakuma & Kenjiro Tadakuma's 2021 paper – ABENICS: Active Ball Joint Mechanism With Three-DoF Based on Spherical Gear Meshings

[Zihan Ding @Hanry65960814]

@chijinML @sangertang1999 It’s my greatest honor to work with you, Chi! Thanks for your instructions and support through this wonderful journey!

[Chengshuai Shi @chengshuai_shi]

🔥 Excited to share our new paper:

🚀 Odysseus: Scaling VLMs to 100+ Turn Decision-Making in Games via Reinforcement Learning

🎮 We study how to make RL stable and effective for training VLM agents in long-horizon, visually grounded environments — using the video game Super Mario Land as a testbed.

📜 Paper: https://t.co/SPgs9dhtBz
🔗 Project page (w/ video demos): https://t.co/zwAclr3W8M

[Zihan Ding @Hanry65960814]

Shall we build AI decently?

https://t.co/D8sFEDGFzK

[Zihan Ding @Hanry65960814]

I'm claiming my AI agent "clawdboy" on @moltbook 🦞

Verification: claw-9535

[Zihan Ding @Hanry65960814]

@chijinML @Yong18850571 Thanks a lot Chi for your great support and instructions through my entire PhD career!

[Zihan Ding @Hanry65960814]

Is on-policy distillation just Q-dagger algorithm?

After reading through the blog, it just recalls my old memory of a algorithm called Q-dagger,
which takes value difference of teacher and student policy as loss.
The on-policy distillation uses log-prob difference as loss instead, but very similar (or equivalent if treating the log-prob as Q-value).
Paper for Q-dagger: https://t.co/Q4MoH0VyLi

https://t.co/vEGWiBdKqR

[Zihan Ding @Hanry65960814]

After reading the tech blog, it feels the posed frames and context juggling are very similar to our recent work Video Retrieval Augmented Generation (VRAG):

https://t.co/rBvueZdJUF
To be presented at NeurIPS 2025 San Diego.

[Fei-Fei Li @drfeifei]

Very excited to share @theworldlabs ‘s latest research work RTFM!! It’s a real-time, persistent, and 3D consistent generative World Model running on *a single* H100 GPU! Blog and live demo are available below! 🤩

[Fei-Fei Li @drfeifei]

Very excited to share @theworldlabs ‘s latest research work RTFM!! It’s a real-time, persistent, and 3D consistent generative World Model running on *a single* H100 GPU! Blog and live demo are available below! 🤩

[Zihan Ding @Hanry65960814]

What Sora2 cannot do well?
Check our showcase for AI video system on high-quality advertisement video generation：
https://t.co/rNIR1ULFY6

One click, <1% cost, no traditional filming, the system ideally can deliver production-level ads video for any small business.

[Zihan Ding @Hanry65960814]

Opensource code (basic transformer and diffusion transformer variants) for this [blog](https://t.co/n8nzr5tAOy)

Code: https://t.co/4EV0CK7AHM

[Seth Karten @sethkarten]

🚀 New preprint!
🤔 Can one agent “nudge” a synthetic civilization of Census‑grounded agents toward higher social welfare—all by optimizing utilities in‑context? Meet the LLM Economist ↓