Huanxuan Liao 廖桓萱

Research statement · Toward Long-Horizon Adaptive Intelligence

My research focuses on efficient and adaptive foundation models, toward the broader goal of long-horizon adaptive intelligence: enabling models to efficiently process and retain expanding contexts, continually acquire and preserve capabilities across evolving tasks, and act reliably in persistent environments under bounded compute, memory, and supervision.

• Efficient Long-Context Modeling: data synthesis, context compression, sparse attention, model architecture, and adaptive visual token allocation, e.g. HyCo₂, SparK, ResAdapt; see also Awesome-LCLM.
• Model Adaptation & Continual Learning: parameter-efficient adaptation, knowledge distillation, and continual learning, e.g. TAGI, AAG, NesyCD, SKIntern, SpaRTA.
• Reliable Long-Horizon Agents: tool use, agent memory, long-horizon evaluation, and reliable decision-making, e.g. CITI, ETAPP, KATE, OPERATE, Distill Yourself, Scientist's Last Exam.