EP Scaling in MoE Inference: Why Prefill and Decode Diverge
Expert parallelism changes weight reuse, token replication, and load imbalance. A simple cost model explains why prefill favors small EP groups while decode benefits from wider ones.
I'm Yanjie. What do I do for a living? I dream.
Research interests
I build self-improving AI systems for automated research.
My work connects algorithm design with scientific discovery, from multimodal generation and ML engineering to biomolecular design.
Learning from generated experience
How can self-generated feedback and synthetic data improve models through post-training?
Scaling computation at test time
How can search and computation scale efficiently for reasoning, design, and optimization?
Learning through interaction
How can computational and experimental feedback enable continual model improvement and scientific discovery?
Latest works
Self-improving AI, automated research, and biomolecular design.
Multimodal Models / Post-training
UniEvo-VL
An On-policy Self-Distillation Training Recipe for Multimodal Model Self-improvement.?
Self-Improving AI / Automated ResearchIn progress
Recursive Self-Improvement for ML Research and Engineering
Implementing and evaluating self-improvement approaches for kernel optimization and ML algorithm development.
Genomic Foundation Model
EVA
A long-context generative RNA foundation model for universal modeling and controllable design.?
Long-horizon Evaluation / Lab-in-the-LoopIn progress
MutaBench
Long-horizon evaluation of scientific agents with lab-in-the-loop feedback.
Alignment/Synthetic data
Proteo-R1 / Proteo-R2
Multi-modal Reasoning for Multi-scale Biomolecule Understanding and Design
Scaling computation at test time
ProteoCraft
AF3-based hallucination with inference-time scaling for binder and TCR designWriting
Notes on machine learning systems, GPU kernels, and efficient inference.
DeepEP Dispatch: Four Designs for Unknown Receive Counts
Compare four DeepEP dispatch designs through buffer ownership, completion signals, and network traffic—and connect those choices to communication latency.
Push or Pull? Communication Design in DeepEP and Mega MoE
The direction of a transfer follows who knows its final address. DeepEP and Mega MoE illustrate how that choice shapes buffering, execution order, and GPU resource use.
