Co-evolution: A "One-to-many" LLM Fine-Tuning Paradigm
Co-Evolve turns a single base LLM into a cooperative population of specialized descendants, using marginal coverage and evolution strategies to improve their complementarity.
I am a third-year undergraduate at Tsinghua University, where I joined the Tsien Hsue-shen Elite Class in Mechanics in 2024. I was a visiting student at the Hong Kong University of Science and Technology (HKUST) from July to August 2026.
My research interests lie in world models, multimodal learning, and embodied intelligence.
I’d love to connect, exchange ideas, and explore opportunities to collaborate. Feel free to reach out!
Learning representations of the world and its dynamics to support prediction and planning.
Connecting vision and language for richer perception, understanding, and reasoning.
Bringing perception, prediction, and action together in agents that interact with the world.
Co-Evolve turns a single base LLM into a cooperative population of specialized descendants, using marginal coverage and evolution strategies to improve their complementarity.
Evaluating proactive multimodal assistants over continuous video streams, with evolving user requests and multi-turn interactions. The work also introduces IPI-Agent to improve when and how models respond.
A benchmark for evaluating how multimodal models reason about interpersonal relationships from videos, including directional relationship judgments and the visual evidence behind them.
Visiting Student
Undergraduate