LLM Agents
From long-horizon reasoning to tools, recovery, and execution in the real world.
Welcome to the homepage of
YuhuaLuo.
Always searching, always learning.
An AI undergraduate at Shanghai Jiao Tong University, currently focusing on LLM agents, multi-agent systems and model collaboration.
BASED INShanghai, China
EDUCATIONSchool of Artificial Intelligence · SJTU
CURRENT FOCUSAgents & collaborative intelligence
01PROFILE
About
Understanding intelligence, in theory and in practice.
Yuhua Luo / silhovette
I am currently an undergraduate student at the School of Artificial Intelligence, Shanghai Jiao Tong University. Alongside my studies, I act as the class monitor and work in the school's Department of Organization.
My research focuses on language-model systems, particularly LLM agents, multi-agent systems, and model collaboration. I am interested in building capable and adaptive agent systems that can reason, coordinate, and act in complex environments. My recent work spans agent planning and execution, multi-agent coordination, and model adaptation.
Collaborative Intelligence
How agents, models, and devices coordinate to solve problems together.
Model Adaptation
How training reshapes model capabilities—and how to extend them without losing what already works.
Selected Honors
Zhiyuan Honor Scholarship, Zhiyuan College, Shanghai Jiao Tong University
Merit Student, Shanghai Jiao Tong University, 2025–2026
Merit Student, Shanghai Jiao Tong University, 2024–2025
Outstanding League Member, Shanghai Jiao Tong University
Timeline
Loading timeline…
02PROJECTS
Selected Work
OctoNest: Adaptive Cross-Device Execution through Stateful Control
ChainLoRA: Geometry-Preserving Task Vector Merging for Continual Learning in LLMs
When Coordination Comes at a Cost: Training-Induced Fragility in Multi-Agent Communication
RELAY: A Heterogeneous Multi-Agent Architecture for Long-Horizon Dynamic Coordination and Multi-Device Control
03RESEARCH
Publications
Flexibility-Reserve-Aware Coordinated Optimal Scheduling of a Park-Level Power–Hydrogen–Electric-Vehicle System for High-Penetration Renewable Energy Accommodation
Coordinated scheduling of power, hydrogen, and electric vehicles with flexibility reserves to support high renewable energy penetration.
OctoNest: Adaptive Cross-Device Execution through Stateful Control
Stateful cross-device orchestration that combines local modality refinement with execution-feedback-driven plan and device reassignment.
ChainLoRA: Geometry-Preserving Task Vector Merging for Continual Learning in LLMs
Replay-free continual learning with chain-updated task vectors, adaptive SVD merging, and geometric separation of shared and task-specific knowledge.
When Coordination Comes at a Cost: Training-Induced Fragility in Multi-Agent Communication
Joint training can make communication fragile. SICT aligns receiver actions across equivalent messages to improve robustness.
What Team Success Hides: Compensation-Masked Capability Loss in Multi-Agent Language Systems
Reveals individual capability loss masked by teammate compensation. RACP selectively rehearses at-risk skills while preserving team utility.
04PERSONAL INTERESTS
Playground
CONTACT
Also at
silhovette@sjtu.edu.cn
where delivery may be unreliable. Please use Gmail for important messages.