More about HKUST
Agent Environment is Key: A Survey towards Scaling Interactive Experience Collection
PhD Qualifying Examination
Title: "Agent Environment is Key: A Survey towards Scaling Interactive
Experience Collection"
by
Mr. Yuchen HUANG
Abstract:
LLM-based agents can autonomously accomplish complex tasks across various
domains. However, to further cultivate capabilities such as adaptive
behavior and long-term decision-making, training on static datasets built
from human-level knowledge is insufficient. These datasets are costly to
construct and lack dynamism, failing to capture the interactive nature of
real-world scenarios. A growing consensus is that agents should instead
interact directly with environments and learn from experience through
reinforcement learning. We formalize this process as the
Generation-Execution-Feedback (GEF) loop, where environments generate
tasks to challenge agents, return observations in response to agents'
actions during task execution, and provide feedback on trajectories,
yielding experience for subsequent training. Under this paradigm,
environments become the primary engines for experience generation, making
their scaling in complexity, dynamism, and interactivity a prerequisite for
collecting high-quality trajectories at scale. In this survey, we adopt an
environment-centric lens and organize existing advances into a unified
taxonomy along the GEF loop. Beyond taxonomy, we further analyze underlying
implementation frameworks, identify open problems, and outline critical
future directions, arguing that the deliberate engineering of interactive
environments will be a decisive frontier in the pursuit of generalist
agents.
Date: Thursday, 30 January 2026
Time: 4:00pm - 6:00pm
Venue: Room 3494
Lift 25/26
Committee Members: Dr. May Fung (Supervisor)
Dr. Yangqiu Song (Chairperson)
Dr. Junxian He