LARK Lab@HKUST (GZ) is led by Dr. Zhijiang Guo. We investigate the fundamental questions of LLMs and Agentic AI to enable rigorous reasoning, robust knowledge integration, efficient algorithms, and generalizable models that achieve superior reliability and scalability.
Research projects
01Agent training
Mosaic
Learning across sessions.
An agent’s work can span many sessions. Its training should, too. Mosaic learns from entire trajectories and the sessions within them, aligning agent training with context compaction and subagent forking.
One rollout. Multiple sessions. Better learning signals.
02Agent inference
SpecAgent
Draft small. Verify smart.
A strong model need not write every step of an agent. SpecAgent lets a small model draft a tool-using trajectory, while a large model reviews it in one pass and repairs only the first wrong step.
Continue from the corrected step Strong reasoning, less generation.
Less expensive generation. Selective, semantic verification.
03Reasoning & memory
QK-Restore
Keep the reasoning. Restore the recall.
Chain-of-thought fine-tuning can teach hybrid language models to reason while disrupting long-range recall. QK-Restore restores pre-training query and key projections, preserving reasoning adaptation without additional training.
