I am interested in building reliable AI agents that can maintain state, use memory, and adapt over long-running interactions. My current research focuses on agent memory, long-running agents, and reliable state-to-action reasoning.
Prospective PhD applicant for Fall 2027
TYDE studies a simple but important question:
When an agent receives new information, what should be remembered, how should existing state be updated, and when should that change the agent's next action?
It is an executable research prototype for selective state updates, scoped memory, and downstream planning in long-running agents. The repository includes an inspectable state-and-memory architecture, diagnostic cases, tests, evaluation scripts, and reproduction instructions.
These repositories document earlier projects across language models, information retrieval, NLP applications, and embedded AI. They are retained as part of my technical development rather than my current primary research focus.
- Mini Llama Training Study — An educational study of small-scale Llama-style model configuration and training using Transformers and TinyStoriesV2.
- Literature Search & BERTopic — A literature-retrieval and topic-analysis prototype combining search, BERTopic, and research-trend exploration.
- Intelligent Consultation Prototype — An undergraduate NLP application exploring AI-assisted medical consultation and information retrieval.
- ESP32 Smart Cat Feeder — An embedded-AI prototype combining lightweight visual recognition with ESP32-CAM-based hardware control.
Additional research and industry projects are described on my personal homepage.
Agent Memory · Long-Running Agents · Self-Evolving Agents · Reliable AI Systems
I am particularly interested in how agents can move from memory → state → action → reusable experience, while remaining reliable as tasks, environments, and interactions evolve over time.
Homepage: yufengliang.me · Email: yufeng.liang13@outlook.com