Zhizhou Sha

CS PhD @UTAustin | Prev CS @Tsinghua_Uni

[Zhizhou Sha @ZhizhouSha]

@GenAI_is_real agree

[Bo Liu @cranialxix]

We had a similar observation before and combined it with the Maglev idea, which becomes Gradlev https://t.co/00qLPwNakp.

In short, you could use one network to predict the costates in parallel, then use prefix scan to simulate token-wise test-time training.

[Zhizhou Sha @ZhizhouSha]

Everything can be calibrated as long as enough money is involved.

[Pingbang Hu 🇹🇼 @PingbangHu]

Prediction market for ICLR: https://t.co/6WGvlzgNSt

yes, this is what we get in 2026 : )

[Lizhang Chen @lzchen_ut]

https://t.co/J4zmqzbOp3

[Zhizhou Sha @ZhizhouSha]

Muse 真的太好用了

我还记得我之前搭建了非常复杂的 harness ，只为了能够让 claude code 和 codex 操纵我的浏览器和我的聊天软件

但是这一切随着muse 的出现，之前一切繁琐的 harness 都没有意义了

我们究竟应该坚守自己的工作流，还是静静等待更好model 和 harness 出现

[Zhizhou Sha @ZhizhouSha]

All people working in agent should definitely check this great work!

[Kevin Wang @KevinWang_111]

Can your AI agent tell when another agent is lying?
Our MindGames report is accepted at NeurIPS! 🎉
944 agent submissions. 76 teams. ~30K games.
Built on TextArena, with game data + top competition agents to test against. Paper and code link below ↓

[Zhizhou Sha @ZhizhouSha]

Exceptional great work 👍

[Max Ryabinin @m_ryabinin]

RL with LLMs is very unstable when training and sampling policies differ. Standard fixes (matching numerics, importance sampling) work around the problem.

We find the root cause of this instability from first principles and propose a way to directly cancel it. Score Centering is competitive and compatible with existing approaches — while simple to implement!

🧵 [1/6]

[Hanqing Zhu @zhu_hanqing666]

@saagnikkk [8/8]
ISO wouldn't exist without this team ❤️

@CongWenyan0320 — my closest friend. @ZhizhouSha, @saagnikkk, Xinyuan Song, David González-Martínez, @XiaoxiaWShirley, @tydsh, @Shiwei_Liu66, David Z. Pan, and Atlas Wang.

[Conference on Parsimony and Learning (CPAL) @CPALconf]

CPAL is so back! We are pleased to announce CPAL 2027 in Tokyo 🇯🇵 March 23-26 https://t.co/vzKPus94xr

[Zhizhou Sha @ZhizhouSha]

@uncle_moon_x 中肯的 社交的本质是价值交换，你得能够拿出别人值得和你交换的东西，社交才会变得有意义

[Zhizhou Sha @ZhizhouSha]

@GenAI_is_real 跟对人走对路

[Bo Liu @cranialxix]

Another way to get parallel training + sequential decoding:

Maglev: Sliding Recurrent Memory
https://t.co/knQhPs5yID

During training, a stronger (e.g., full-attention) teacher provides recurrent memory states, allowing the student (e.g., sliding window attention) to predict all tokens in parallel. The student learns to reproduce those states itself, closing the train–inference gap.

At inference, the teacher disappears: the student rolls its memory forward sequentially. Interestingly, one could have student and teacher share most of the parameters : )

[SystemPromptIndex @aisystemprompt]

https://t.co/vfc7H6ZZdQ

[Zhizhou Sha @ZhizhouSha]

@Sunqi_Fan congrats!

[Zhizhou Sha @ZhizhouSha]

Grateful to all collaborators 🙏 Thanks for reading!

Authors: Xu Yang @xyyy6688, Zhizhou Sha, Junbo Li @ljb121002, Jian Yu, Yifan Sun, Matthew Zhao, Jinrui Fang, Xinyue Guo, Yining Wu, Xu Hu @CamillaHu2002, Yifu Luo, Qiang Liu, Zhangyang Wang @VITAGroupUT

Affiliation: @UTAustin

[Zhizhou Sha @ZhizhouSha]

Can you boost your AI review scores without changing any of the science?

Yes, no hidden text, no prompt injection. Just repackage how the paper reads(figures, equations, numbers all fixed). We call it adversarial repackaging.

→75.1% attack success, +1.21/10 mean score gain🧵👇

[Zhizhou Sha @ZhizhouSha]

Open question: if presentation alone can move AI review scores, what's the right fix? Better reviewer prompts, ensembles, or keeping humans in the loop?

We'd love to hear how you'd make AI reviewers robust to this.