Huanxuan Liao 廖桓萱

Research statement

I work on making large language models and multimodal LLMs practical under budget: long inputs, dense visual streams, and interactive agents should not require unbounded compute or memory. My research centers on three coupled directions— Long-Context & Multimodal Efficiency (input-side budget allocation, hybrid context / KV compression; e.g. ResAdapt, HyCo2, SparK), Efficient Reasoning & Policy Optimization (activating and transferring parametric / symbolic knowledge under cost-aware objectives; e.g. TAGI, SHIP, SpaRTA, BuPO / CAPO), and Agentic Systems & Tool Learning (tool use, evaluation, and memory for reliable agents; e.g. CITI, ETAPP, KATE, Distill Yourself). Across these lines I aim for methods that stay compatible with real inference stacks and learn stably from interaction and feedback.