News
-
May 01, 2025
Two papers will be presented at ICML 2025. Congrats to Guanzheng Chen, Qilong Feng, Keyu Duan, Yiran Zhao and our collaborators!
Unnatural Languages Are Not Bugs but Features for LLMs
Long-Context Inference with Retrieval-Augmented Speculative Decoding (Spotlight) -
January 23, 2025
One paper will be presented at NAACL 2025. Congrats to Xiangyan Liu, and our collaborators!
CodexGraph: Bridging Large Language Models and Code Repositories via Code Graph Databases -
January 23, 2025
Three papers will be presented at ICLR 2025. Congrats to Jinjie Ni, Yiran Zhao, Guanzheng Chen, and our collaborators!
MixEval-X: Any-to-Any Evaluations from Real-World Data Mixtures (Spotlight)
Identifying and Tuning Safety Neurons in Large Language Models
LongPO: Long Context Self-Evolution of Large Language Models through Short-to-Long Preference Optimization
Featured Research
Safety Neuron
We propose a neuron detection method to identify and tune safety neurons—less than 1% of model parameters—using SN-Tune, which enhances safety without affecting general model performance.
Noisy Student
The first method that uses extra unlabeled data to achieve state-of-the-art on ImageNet. Also employed in AlphaFold 2 (Nobel Prize 2024), Google Search and other state-of-the-art AI systems.
People
Faculty
PhD Students
Hannah Brown
Esther Gan
Yiran Zhao
Direction
What’s a human experience? Michael had this question one day. The initial answers were all around achievements. But a complementary perspective later arose as “experiencing rock music played in my car, trees outside and everything as it is”.
We work on democratizing effective and efficient AI models and systems in the current paradigm. We are also interested in directions which might fundamentally change how AIs experience and relate to the world.
So we named our lab Tree and Rock AI Lab (TRAIL). Enjoy the trail with us, as a reader, as a member or as a supporter!
Publications
*: equal contribution, †: equal advising
