
@AWS | prev @WisconsinCS @RiceCompSci, @Google, @ThirdAILab (acq. by @ServiceNow), avid fan of @Uruguay
Joined February 2019
- Tweets42
- Following1.3K
- Followers268
- Likes114
Minghao Yan retweeted
🤔 Is score centering (SC) all we need for stable async RL?
People often stack truncated importance sampling (TIS) or KL penalties on top of SC. In a fun side project, we tested how much they help.
We found that SC alone often matched on-policy. But when it fell short, adding more corrections doesn't necessarily help. 🧵
Check it out!!
I’ll be at @COLM_conf this week (Oct 5–10), presenting work from my internship at ByteDance Seed:
“Towards Full Pipeline FP8 Reinforcement Learning for LLMs”
arxiv.org/abs/2609.22870
Happy to chat about RL infra, agentic RL, and continual learning. Come say hi if you’re around :)
222
🚀🚀🚀
Scientific American has named ODBI’s Kaiyi Jiang to its inaugural list of Young American Scientists. Congrats, @idmjky!
Jiang will join ODBI next month. Read more on the ODBI site: bioengineering.princeton.edu…
2
478
Minghao Yan retweeted
1/8) Excited to share PR²: Predictive Routing Replay for MoE-Based LLM RL! 🎉
This is also the first paper from our research group at @RutgersCS @RutgersU.
While MoE LLMs scale remarkably well, RL training exposes a hidden source of instability: routing.
Paper: arxiv.org/abs/2606.00395
A special shout-out to my talented PhD student, @DaizeDongCS, for leading this project and driving many of its key ideas and experiments.
🚀 Thrilled to introduce PACEvolve: Enabling Long-Horizon Progress-Aware Consistent Evolution.
We show how to push LLM self-evolution beyond short, unstable improvements and into consistent, long-horizon gains. 🧵👇
This work was done during my internship at Google and would not have been possible without my mentors and collaborators across Google and DeepMind. Kudos to everyone involved!
Paper: arxiv.org/pdf/2601.10657
Code drop coming soon, stay tuned!
1
3
250
Code drop here:
Our code is finally out at github.com/Google/pacevolve.
Run it with your favorite task and see if you can push the scientific frontier!
235
Our code is finally out at github.com/Google/pacevolve.
Run it with your favorite task and see if you can push the scientific frontier!
I will be presenting our work, Decoding Speculative Decoding, at @naacl tomorrow. We identified the performance bottleneck in speculative decoding to be draft model depth and demonstrated low correlation between language modeling performance and token acceptance rate.
We showed that even for LLaMA 3.1 models, which were directly distilled from the target model, our lightweight method can still improve throughput by 52% with less than 10 minutes of distillation.
1
297















