Mengdi Wang
5 indexed papers
Publications per year
Top categories
Frequent co-authors
Research Timeline
The paper introduces Science Earth, a planet-scale scientific runtime that enables diverse, siloed AI capabilities to connect and collaborate dynamically, demonstrating that scientific discovery can become a distributed, self-correcting process.
The paper introduces Science Earth, a planet-scale scientific runtime that enables diverse, siloed AI capabilities to connect and collaborate dynamically, demonstrating that scientific discovery can become a distributed, self-correcting process.
EEVEE is a multi-dataset test-time prompt learning framework for LLM agents.
This paper introduces Skill Entropy, a measure of difficulty for switching between skills in long-horizon reasoning tasks, and proposes Skill^2-Bench, a benchmark for evaluating cross-skill long-horizon tasks. The authors also present Skill-Entropy RL, a reinforce learning framework using skill entropy as a training signal.
This paper proposes PFM, a dual-view memory system for NPU-PIM systems that decouples physical data layout from accessor-visible logical views, reducing bandwidth underutilization and performance loss.
Papers
Rethinking Unified Memory for NPU-PIM Systems: Dual-View Memory for Dynamic Inference of LLM
Shixin Zhao, Lian Liu, Tianhua Han, Mengdi Wang +2 more
This paper proposes PFM, a dual-view memory system for NPU-PIM systems that decouples physical data layout from accessor-visible logical views, reducing bandwidth underutilization and performance loss…