Huanxuan Liao 廖桓萱

Research statement · Long-Horizon Adaptive Intelligence

I study foundation models for long-horizon adaptive intelligence: models that use growing contexts effectively, acquire and retain reusable capabilities across evolving tasks, and act reliably over extended interactions.

• Long-Context Data & Model Co-Design: long-context data curation and synthesis, length- and capability-aware mixtures, scalable pretraining and mid-training, efficient attention, KV-cache optimization, and architectures to jointly improve capability, scalability, and efficiency (Awesome-LCLM, HyCo₂, SparK, ResAdapt, AllocEmbed).
• Knowledge & Continual Adaptation: mechanisms for acquiring, transferring, internalizing, and updating knowledge and capabilities across evolving tasks, while balancing plasticity, retention, and generalization (TAGI, SHIP, DyPRAG, AAG, NesyCD, SKIntern, SpaRTA).
• Agent Learning & Self-Improvement: persistent memory, long-horizon decision-making, and learning from interaction and feedback, toward agents that verifiably refine their memory, policies, and learning procedures (CITI, KATE, BuPO, Tool-RL-Box, OPERATE, Distill Yourself, Scientist's Last Exam).