Xiangning Chen (@XiangningChen) / X

Xiangning Chen

145 posts

Xiangning Chen

@XiangningChen

Research @Meta | Previously: @OpenAI @GoogleDeepMind @UCLA @Tsinghua_Uni | Opinions are my own

Joined May 2017

Pinned
Xiangning Chen
@XiangningChen
Apr 8
Very satisfying to see the beautiful RL scaling curves. It's a smart and token efficient model. Amazing team effort! 🚀
Hongyu Ren
@ren_hongyu
Apr 8
Check out Muse Spark, our first milestone in the quest for personal superintelligence! Scaling this with the team has been a total blast. Give it a spin and let us know what you think! 🥑
7.8K
Xiangning Chen
@XiangningChen
Nov 12, 2025
Proud to see this out in the world. Spent quite some time pushing on instruction following and it is exciting to see it land and get recognized. Huge team effort and grateful for everyone who made GPT 5.1 shine!
Sam Altman
@sama
Nov 12, 2025
GPT-5.1 is out! It's a nice upgrade. I particularly like the improvements in instruction following, and the adaptive thinking. The intelligence and style improvements are good too.
146K
Xiangning Chen
@XiangningChen
Jun 13, 2023
I've just been told that MPT-7B is trained by our Lion optimizer (arxiv.org/abs/2302.06675). Lion has also been successfully deployed in production systems such as Google search ads CTR model. Glad to see that our work has real-world production impact! mosaicml.com/blog/mpt-7b
arxiv.org
Symbolic Discovery of Optimization Algorithms
We present a method to formulate algorithm discovery as program search, and apply it to discover optimization algorithms for deep neural network training. We leverage efficient search techniques...
95K
Xiangning Chen
@XiangningChen
Feb 16, 2023
Symbolic program search discovers an effective optimizer that only tracks momentum, Lion🦁. It gains 88.3% 0-shot and 91.1% finetuning ImageNet accuracy, and up to 5x, 2.3x, and 2x training efficiency on ViT, diffusion models, and LM, compared to Adam. arxiv.org/abs/2302.06675
19K
Xiangning Chen
@XiangningChen
Aug 7, 2025
Replying to @aidan_mclau
It’s an honor working with you Aiden! 🐐
6.2K
Xiangning Chen
@XiangningChen
May 24, 2023
Thanks for the great work on the exciting 2nd-order optimizer! I found Lion has the same lambda as AdamW while with a 5x smaller lr. Sec 5 of our paper shows that it should be increased to maintain a similar strength. Would be interested to see the results of tuned Lion.
Tengyu Ma
@tengyuma
May 24, 2023
Adam, a 9-yr old optimizer, is the go-to for training LLMs (eg, GPT-3, OPT, LLAMA). Introducing Sophia, a new optimizer that is 2x faster than Adam on LLMs. Just a few more lines of code could cut your costs from $2M to $1M (if scaling laws hold). arxiv.org/abs/2305.14342 🧵⬇️
29K
Xiangning Chen
@XiangningChen
Jun 1, 2023
🔒Detecting AI-written texts is crucial in managing safety and ethical concerns. 💥We show that protected LLMs alone can bypass current detectors like DetectGPT, Watermarking, and OpenAI's classifier without relying on any unprotected paraphrasing model.
arxiv.org
Red Teaming Language Model Detectors with Language Models
The prevalence and strong capability of large language models (LLMs) present significant safety and ethical risks if exploited by malicious users. To prevent the potentially deceptive usage of...
11K
Xiangning Chen
@XiangningChen
Jun 13, 2023
Replying to @typedfemale
Thanks for the reminder. This information is internally approved and our latest arxiv version contains it already.
1.3K
Xiangning Chen
@XiangningChen
Jun 13, 2023
Replying to @XiangningChen
"We also train our MPT models with the Lion optimizer rather than AdamW, which provides stable update magnitudes and cuts optimizer state memory in half."
3.1K
Xiangning Chen
@XiangningChen
Aug 7, 2025
Excited to bring GPT-5 to you! Grateful to work with a brilliant team and contribute to several workstreams making GPT-5 and GPT-5 thinking more steerable, conversational, intelligent, safe, and full of personality!
OpenAI
@OpenAI
Aug 7, 2025
GPT-5 is here. Rolling out to everyone starting today. openai.com/gpt-5/
00:00
2.7K
Xiangning Chen
@XiangningChen
Nov 13, 2025
Replying to @sama
Thank you, Sam! Really appreciate your support.
2.5K
Xiangning Chen
@XiangningChen
Aug 11, 2025
From 49th to 98th percentile in IOI in a single year. Insane progress.
Sheryl Hsu
@SherylHsu02
Aug 11, 2025
1/n I’m thrilled to share that our @OpenAI reasoning system scored high enough to achieve gold 🥇🥇 in one of the world’s top programming competitions - the 2025 International Olympiad in Informatics (IOI) - placing first among AI participants! 👨‍💻👨‍💻
1.3K
Xiangning Chen
@XiangningChen
Aug 7, 2025
Steerability matters not only for chat, but also for coding and many other real-world use cases. Excited to see our personality steerability effort also helps make GPT-5 and GPT-5 thinking more steerable on Cursor. Congrats to the team! @XinweiS @ZhengWu777 @qinghual2020
Cursor
@cursor_ai
Aug 7, 2025
GPT-5 is now available in Cursor. It’s the most intelligent coding model our team has tested. We're launching it for free for the time being. Enjoy!
1.7K
Xiangning Chen
@XiangningChen
Oct 28, 2025
Replying to @zephyr_z9
This table is from Qwen3 technical report bro…
984