I am an MSc candidate in Software Engineering at Xiamen University (Sep 2025–present). My work sits at the intersection of long-horizon language agents, context-aware AI systems, and reliable software engineering: I study how an intelligent system selects, organizes, updates, and reuses information over extended tasks, and I turn those questions into traceable and reproducible research software.

Research agenda

Context management for long-horizon language agents. When an agent works across many steps, the problem is not simply whether a model has a larger context window. The system must decide which observations to retain, which information to compress, when to retrieve earlier evidence, and how to keep the resulting context relevant to the current task. My work explores dynamic long-context management, adaptive working and long-term memory, multi-source memory conflict detection and resolution, and the construction of minimal but sufficient task context from development history, code, and documentation.

Learning and memory-update mechanisms. I am interested in how an agent can improve through interaction rather than merely accumulate text. This includes using task history as a learning signal, updating memory when new evidence arrives, distinguishing stable knowledge from temporary working state, and evaluating whether a memory update actually improves later decisions. In this direction, evaluation is part of the learning mechanism: an agent should expose what it remembered, why it used that information, and whether the update changed its behavior.

Language intelligence and conceptual transfer. My deeper interest has always been language itself. I often think of a language model as a “ghost of language”: it does not experience the world as a person does, yet it can reorganize concepts, relations, and action traces from the language left behind by people. In Concept2Fable, I study how a concept or mechanism can move between domains without losing its structural constraints, and how that transfer can be evaluated through evidence alignment and structural analogy rather than surface fluency alone.

Research trajectory

• 2025–2026 — Affective computing and multimodal understanding: began with emotion-related questions in affective computing, then worked on multimodal aspect-based sentiment analysis through literature review, model reproduction, experimental comparison, and optimal-transport-based cross-modal alignment. I also developed ContextLens, a mixed-reality learning system that organizes questions, video frames and captions, book content, and scene context into a traceable Context Tree for grounded multimodal LLM responses.
• 2026 — Language intelligence and conceptual transfer: designed the knowledge-graph and LLM pipeline in Concept2Fable for mechanism-constrained statements, educational narratives, and cross-domain fables.
• 2026–present — Long-horizon agents: investigate adaptive working-memory compression, long-term memory environments, memory conflict resolution, and traceable context construction for coding-agent tasks.

Current engineering focus

I work primarily with Python and PyTorch, alongside Graph RAG, tool use, coding agents, Linux, Docker, pytest, and reproducible experiment workflows. My engineering questions are concrete: what information is retained or discarded at each step, how a memory update can be inspected and reversed, which evidence supports a generated answer or action, and whether an evaluation task detects the behavior that actually matters. I also care about turning the surrounding code, experiments, notes, and documentation into a reusable working context rather than leaving them as disconnected artifacts.

I am building LLM.learn, a public curriculum and knowledge system for learning LLMs, agents, memory, and AI infrastructure. The goal is to organize public courses into a coherent path of source materials, transcripts, structured notes, technical explanations, and reproducible exercises.

Beyond research

Outside research, I read fiction and books in the humanities and social sciences, and I take related courses when they help me understand language, concepts, and people from another angle. I also enjoy basketball, football, and endurance running, including marathon-distance races.