Scholar
Li Lyna Zhang
Google Scholar ID: -_ItfAoAAAAJ
Microsoft Research Asia
Citations & Impact
Citations
3,465
H-index
18
i10-index
23
Publications
20
Co-authors
12
list available
Contact
Publications
Resume (English only)
Academic Achievements
- August 2024: Introduced rStar, a self-play mutual reasoning approach that significantly boosts reasoning capabilities of Small Language Models (SLMs) during inference—e.g., improved GSM8K accuracy of LLaMA2-7B from 12.51% to 63.91%
- rStar has been recommended as a key technique in OAI-o1-like approaches
- August 2024: Contributed to the release of Phi3.5-128k LLMs with significant improvements in LongRoPE for recovering short-context performance after context window extension
- July 2024: Released LongRoPE-related work
- Work featured on Hugging Face Daily Papers and Jiqizhixin (Machine Heart)
Background
- Currently a Principal Researcher in the Systems and Networking Group at Microsoft Research Asia (MSRA)
- Broad research interests in AI algorithms
- Since joining MSRA, focused on novel algorithms for improving AI inference efficiency, including: (1) compression for pre-trained Transformer models and LLMs; (2) hardware-aware Neural Architecture Search (NAS) for edge AI
- Recently deeply engaged in exploring cutting-edge research problems in Large Language Models (LLMs) and Artificial General Intelligence (AGI)
- Actively working on topics such as long-context LLMs and LLM self-play
