How a chat context is kept as a list of role-tagged messages, tried on Qwen3-1.7B.
About
I study how language models hold on to context across long, multi-turn conversations — and how to run them where the user is.
I’m Hojun Lee (이호준), an incoming graduate student working on AI agents. Before that I built LLM tools as a developer, some of which live at foxisdogmung.dev.
Interests
- Context management in multi-turn LLM conversations
- On-device LLM inference on NPUs
- Scheduling agentic workloads on heterogeneous SoCs
Recent writing
Paper notes: measuring the gap between single-turn evals and real multi-turn use by sharding instructions.