From Agentic AI toward Reliable Visual World Modeling for Long-Horizon Decision Making

PhD Qualifying Examination


Title: "From Agentic AI toward Reliable Visual World Modeling for Long-Horizon
Decision Making"

by

Mr. Meng CHU


Abstract:

Agentic AI is moving from single-turn reasoning toward long-horizon visual
decision making, where failures often stem from weak action-conditioned
prediction rather than weak language generation. This survey examines progress
from 2024 to early 2026 through the lens of visual world modeling for long-
horizon reliability. We organize the literature around two system components:
a Simulator for decision-time visual rehearsal and an Evolver for converting
deployment evidence into persistent, verifiable improvements.

We propose a unified taxonomy across representation, dynamics, and control
interfaces, and analyze recurring failure modes—including compounding error,
state aliasing, exploitability, and out-of-distribution miscalibration—that
limit long-horizon performance. We further summarize decision-centric
evaluation protocols for visual settings, emphasizing action-response
consistency, replayability, tail-risk reporting, and benchmark robustness.

Building on this analysis, we outline a thesis-oriented research agenda
centered on controllable, calibratable visual world models that improve
reliability, recovery, and safety under distribution shift. We argue that
progress in visual agent systems should be measured not by perceptual realism
alone, but by whether visual predictions systematically improve decision
quality and robustness in interactive settings.


Date:                   Tuesday, 17 March 2026

Time:                   3:00pm - 5:00pm

Venue:                  Room 2132C
                        Lift 22

Committee Members:      Prof. Jiaya Jia (Supervisor)
                        Dr. Qifeng Chen (Chairperson)
                        Dr. Long Chen