I’m an undergraduate at the Gaoling School of Artificial Intelligence, Renmin University of China.
My current research interests lie in Multimodal Reasoning, Efficient Multimodal Models, and Video Generation. In particular, I am interested in understanding and reducing computational redundancy in multimodal models through KV cache compression, token pruning, sparse attention, and adaptive caching.
I also have experience in Visual SLAM and robotic perception, which has shaped my interest in building efficient and robust intelligent systems that can perceive, reason about, and interact with the real world.
Connect with me:
-000000?style=for-the-badge&logo=x&logoColor=white){kind=icon}
-
Constraint-Guided Prompting and Semantic-Aware Evaluation for LLM-Based ABSA
— Muzhi Li, Tiancheng Xing, Yuheng Wang, ICIC 2026 · Springer LNAI
- Introduces Constraint-Guided Prompting (CGP) for more reliable structured extraction with large language models and Sem-F1, a semantic-aware evaluation protocol for Aspect-Based Sentiment Analysis.
-
- A source-grounded character conversational AI that uses RAG over original game scripts to preserve character knowledge and persona. It combines FAISS-based semantic retrieval, pluggable local/cloud LLM backends, multi-turn session memory, real-time SSE streaming, and an immersive Vue 3 visual-novel interface.
-
- A C++20 monocular visual SLAM and dense reconstruction system inspired by ORB-SLAM3. It combines geometric tracking, local bundle adjustment, Sim(3) loop closure, and pose-graph optimization with learned monocular depth, using sparse SLAM landmarks and multi-view consistency to reconstruct a colored dense 3D map.
- Multimodal Reasoning: Vision-Language Models · Long-Horizon Reasoning · Visual Information Flow
- Efficient Multimodal Inference: KV Cache Compression & Eviction · Token Pruning · Sparse Attention · Dynamic Budget Allocation
- Video Generation: Diffusion Models · Flow Matching · Video DiT · Autoregressive / Chunk-wise Generation · Feature & KV Caching
- Embodied Intelligence: Vision-Language-Action Models · World Models · Efficient Embodied Reasoning
