Real-time intelligence · Embodied agents · Interactive AI
Research interests: post-training and inference infrastructure; on-device embodied real-time agents, with faster structured decisions, scalable embodied agents, and on-device AI-native applications.
Ph.D. student at Shanghai Innovation Institute / Shanghai Jiao Tong University, advised by Prof. Zengfeng Huang.
My work spans LLM post-training, inference systems, and on-device products. The recurring question is: when an intelligent model must perceive state, decide, and affect an interface or physical system within tens of milliseconds, how should the model, action space, and execution stack be designed together?
Research, systems, and product
ICML 2026 · First author
Parallel Decoding for Real-Time LLM Function Calling
After one shared prefill, function names and arguments are generated concurrently across parallel heads, removing the serial structured-decision bottleneck in real-time agents. It delivers 3–6× typical end-to-end speedup and up to 9.6×, with RT-Qwen3-4B reaching 61.2 ms P50 on an RTX 4090.
Ongoing research · Embodied Agent-as-Policy
Typed tool calls as a scalable embodied action space
SimpleToolVLA treats tool calling as the high-level interface between an LLM / VLM and controllers, VLA skills, or world-action models. Dynamic schemas define typed, composable, and verifiable semantic actions. The current work tests two falsifiable hypotheses: whether simulation and synthetic tool trajectories scale transferable policies, and whether the resulting edge loop can reliably meet physical deadlines.
Startup · AI-OC-native interactive content
Personal OC for the Interactive Internet
SimpleLove explores persistent AI-OC identity, cross-experience memory, and prompt-to-interactive-content.
One side is a content experience shaped by an OC, memory, and preference; the other is a creation platform
where agents help modify, test, and remix runnable experiences. The long-term goal is for the same OC to
grow from a character inside content into a personal agent on the user's devices.
In short: a “TikTok” for OC mini-apps, driven by coding agents and AIGC.
miHoYoIntern
Shanghai Innovation Institute / Shanghai Jiao Tong UniversityPh.D. student · post-training and infra · Advisor: Prof. Zengfeng Huang
Shanghai Jiao Tong UniversityB.S. in Chemistry · Computational chemistry
Open invitation
SimpleLove is looking for people who share a long-term view of on-device intelligence, AI-native interactive content, and character intelligence. I welcome conversations with investors, founders, researchers, engineers, and creators about funding, collaboration, or joining the team.