Xinye Li

Interest: LLM | Knowledge | Agent | World Models

[Xinye Li @vclee8]

@_akhaliq

[Xinye Li @vclee8]

@_akhaliq I've just open-sourced an adjacent work — ForgeWM, which applies the same paradigm to discrete keyboard/mouse control on a Minecraft world model (Matrix-Game 2 lineage), 8 GPUs, fully open data/weights/code. Different cell of the same space.

[Xinye Li @vclee8]

8/ #WorldModels

Thanks for reading this far. 🙏

If you've worked on world models, video diffusion, or distillation
— I'd love to hear what you'd want from a project like this.
Always happy to chat.
↓ QR for the repo if it's easier.

[Xinye Li @vclee8]

We just open-sourced ForgeWM — a fully reproducible recipe for training a playable Minecraft world model on 8 GPUs. Keyboard + mouse control. Causal Forcing distillation. 4-step real-time inference.
Code, weights, data — all open. 🧵 (1/7)

[Xinye Li @vclee8]

7/ Code, weights, data, project page:
🔗 https://t.co/L4wqEvk43i
🌐 https://t.co/5NwX1QKpii
🤗 https://t.co/RpADSmnf8W
I'm an early researcher just getting into world models —
feedback, PRs all welcome.

Built standing on the shoulders of MG2, GameFactory,
Causal Forcing 🙏

[Mathieu @miniapeur]

Every 5 neurips/icml papers PhD student interviewing for the same internships at DeepMind like

[Xinye Li @vclee8]

@lihanc02 @sheriyuo Is there any reason? US law has more restrictions? or is it an overall management issue with US PhD programs (like TA stuff?)

[Xinye Li @vclee8]

@sheriyuo Gaoling不能实习吗？似乎在外面实习挺多的？（只是听说

[Xinye Li @vclee8]

Maybe a new proxy to examine the dataset contamination? Interesting

[Anthropic @AnthropicAI]

New Anthropic research: Natural Language Autoencoders.

Models like Claude talk in words but think in numbers. The numbers—called activations—encode Claude’s thoughts, but not in a language we can read.

Here, we train Claude to translate its activations into human-readable text.

[Arun @hiarun02]

Claude Code 4.7 is insane.

i know literally NOTHING about coding. ZERO. and i just built 3 fully functioning web apps in 30 minutes.

http://localhost:3000/
http://localhost:8000/
http://localhost:5000/

check it out.

[Xinye Li @vclee8]

下一轮进展在哪？
1. 构造"局部最优"的合成教师。
2. 不追求最聪明的老师，而是针对当前学生、当前 checkpoint、每步最有帮助的老师。
3. 目标是在"蒸馏 ↔ RL"之间画一条连续曲线：不依赖真实同族教师，每一点都算力最优。

这是作者押注的开放问题。（8/8)

[Xinye Li @vclee8]

读完 Will Brown 新博客《On SFT, RL, and on-policy distillation》，信息密度极高，也让我重新把后训练的这些算法串了一遍。

一句话总结：

后训练的真正框架不是"SFT or RL"，而是"选一个 KL 预算 β，再为它找一个局部最优的教师"。

这是我的阅读笔记👇

[will brown @willccbb]

https://t.co/gCIFKAjB0Z

[Xinye Li @vclee8]

所有方法都是同一个元算法的特例

三个旋钮：
· α：采样多 on-policy
· λ：信号是教师 KL 还是 outcome reward
· π_T：教师是谁、在什么条件下

SFT / RL / OPD / OPSD 只是这个空间的不同角点。

但作者提醒：不要真去插值 α、λ，真正重要的轴是 KL 预算 β 和教师选择。