More about HKUST
From Self-Aware to Human-Aware: A Survey on Trustworthy LLM Agents
PhD Qualifying Examination
Title: "From Self-Aware to Human-Aware: A Survey on Trustworthy LLM
Agents"
by
Mr. Zhitao HE
Abstract:
Large language model (LLM) agents are increasingly deployed as interactive
systems that reason, use tools, communicate with users, and act in
consequential environments. This shift raises a central question for
trustworthy AI: what must an agent be aware of before it can act reliably on
behalf of humans? This survey organizes recent work around an
awareness-centered taxonomy of trustworthy LLM agents. We review three
complementary forms of awareness: self-awareness, where agents recognize their
own knowledge boundaries, uncertainty, capability limits, and
self-verification mechanisms; environment-awareness, where agents ground
decisions in external constraints, tools, domains, and verifiable
consequences; and human-awareness, where agents model mental states, user
intent, preferences, collaboration, and group dynamics. Rather than treating
trustworthiness as a flat collection of safety, robustness, and alignment
techniques, we argue that these dimensions can be understood as failures or
successes of agent awareness. For each form of awareness, we organize the
relevant methods, benchmarks, and training approaches within the corresponding
section. We conclude with open challenges in distinguishing genuine awareness
from performative claims, scaling evaluation to realistic environments, and
balancing helpful autonomy with safety-critical boundaries.
Date: Monday, 15 June 2026
Time: 5:00pm - 6:00pm
Venue: Room 3494
Lift 25/26
Committee Members: Prof. Siu-Wing Cheng (Supervisor)
Dr. May Fung (Co-Supervisor)
Prof. James Kwok (Chairperson)
Dr. Sirui Han (EMIA)