Seeing What Matters: Building Reliable Multimodal Intelligence

PhD Thesis Proposal Defence


Title: "Seeing What Matters: Building Reliable Multimodal Intelligence"

by

Mr. Kaican LI


Abstract:

Multimodal foundation models have achieved remarkable capabilities, yet they 
remain unreliable when they rely on spurious correlations, overlook critical 
visual evidence, or become overwhelmed by complex and information-dense 
inputs. This thesis investigates how multimodal systems can better identify 
and act on the information that matters. First, it develops learning 
objectives that encourage models to rely on semantically consistent and 
robust features, improving generalization under distribution shift. It then 
moves from passive evidence use to active perception, introducing agents that 
dynamically search for relevant visual information and selectively acquire 
higher-resolution evidence during reasoning. Together, these works establish 
a progression from learning reliable representations to building multimodal 
agents that actively control their perception, with the broader goal of 
creating intelligent systems that know what to trust, where to look, and when 
to examine evidence more closely.


Date:                   Tuesday, 13 October 2026

Time:                   3:00pm - 5:00pm

Venue:                  Room 5504
                        Lifts 25/26

Committee Members:      Prof. Nevin Zhang (Supervisor)
                        Prof. Gary Chan (Chairperson)
                        Dr. Yangqiu Song