Pengyu Zhao leads large-language-model research and engineering at MiniMax and serves as an executive director. His co-authored research includes long-context foundation models and verifiable logical-reasoning data for reinforcement learning.[1][2][3]
Start with SynLogic code and data →
Career
At Hulu’s Beijing technology organization, Zhao worked on recommendation algorithms.[1]
Contributions
Generate reasoning problems with checkable answers
SynLogic supplies generators and rule-based verifiers for 35 logical-reasoning tasks. Difficulty and dataset size can be adjusted, allowing reinforcement learning to use answers that can be checked rather than an additional model’s opinion. A puzzle is generated, expressed as instructions and paired with its task-specific verifier. Grid size can control a Sudoku instance’s difficulty, while another puzzle needs its own rules. Zhao co-authored this work with the MiniMax-led research team.[2]
Extend context through model architecture
MiniMax-01 combines lightning attention with a mixture of experts and systems work for training and inference. Its hybrid architecture places one softmax-attention block after every seven lightning-attention blocks. Lightning attention provides an efficient linear-attention implementation for long contexts; the report also describes overlapping computation and communication. The team reports million-token training contexts and longer inference extrapolation. Zhao is named among its contributors.[3]
Tools, writing & talks
Tools
Start here
SynLogic code and data
github.comGenerators and verifiers for adjustable logical-reasoning training data.
Start with the task examples and one generator. Reading requires no training setup; generating data needs Python dependencies, and reinforcement learning needs a separate training environment.
Writing, research & talks
- MiniMax-01arxiv.org
A team report on lightning attention, mixture-of-experts models and long-context systems.
Named team contributor[3]