About
Mighten Dai - A tech enthusiast exploring AI systems, Large Language Models, and Software Engineering.
Hi, I’m Mighten Dai, an Applied Scientist focusing on Large Language Models, Multimodal AI, Reinforcement Learning, and AI Agent systems.
I am currently pursuing my M.S. in Software Engineering at the University of Science and Technology of China. My research interests lie at the intersection of foundation models and real-world intelligent systems — especially how we can improve the reasoning, reliability, and usability of LLMs through better training strategies, reinforcement learning, and system design.
My work spans both research and engineering. I am interested in building AI systems that are not only capable of reasoning, but also reliable, efficient, and deployable in real-world scenarios.
My recent research focuses on post-training and reasoning enhancement for Large Language Models and Vision-Language Models.
My interests include:
- Reinforcement Learning for LLM alignment and reasoning
- Multimodal Large Language Models (MLLMs)
- Supervised Fine-Tuning (SFT) and RL-based post-training
- GRPO-based optimization
- Efficient model inference and deployment
I explore how structured representation, reward design, and training strategies can unlock stronger reasoning abilities in Large Language Models.
My recent work includes improving VLM reasoning for chart-to-code generation through structured representations and reinforcement learning, enhancing generation reliability and structural accuracy.
I am also interested in building more trustworthy AI systems, such as improving anomaly detection and reasoning capabilities through Chain-of-Thought supervision and reinforcement learning.
Beyond research, I enjoy turning AI ideas into practical systems.
During my internship, I worked on Retrieval-Augumented Generation (RAG) and AI Agent systems.
My engineering experience includes:
- Designing multi-stage retrieval pipelines for enterprise RAG systems
- Improving retrieval quality and QA accuracy
- Building AI Agent tool execution infrastructure
- Developing sandbox-based execution services with Docker and Flask
- Optimizing real-time LLM interaction systems and reducing latency
I am especially interested in the transition from model capability to product capability — transforming research ideas into reliable AI applications.
I believe open source is one of the best ways to learn, collaborate, and contribute to the AI community.
I contribute to projects related to LLM infrastructure and cloud-native systems, including:
- verl — an open-source LLM post-training framework for reinforcement learning workflows
- Apache DolphinScheduler — a distributed workflow orchestration platform
- Research projects around multimodal reasoning and AI systems
My contributions include:
- Improving training stability for multimodal RL workflows
- Fixing framework-level issues
- Extending system capabilities
- Improving usability and reliability of AI infrastructure
You can find my projects and contributions on my GitHub:
This blog is where I share my thoughts and experiments around:
- Large Language Model training
- Reinforcement Learning for AI
- Information Retrieval, RAG, and AI Agent
- AI infrastructure
- Open-source engineering
- Cloud Computing
- Lessons learned from building intelligent systems
I enjoy understanding AI systems from first principles — from algorithms and model architectures to the engineering details behind scalable intelligent applications.
Building intelligent systems requires both scientific curiosity and engineering discipline.
I believe the future of AI will come from the combination of:
- rigorous research
- careful engineering
- and continuous exploration
🚀🚀 PER ASPER AD ASTRA 🚀🚀