More about HKUST
Grounding Judgement in Structure and Evidence: From Structured Generation to Agentic Reasoning
PhD Thesis Proposal Defence
Title: "Grounding Judgement in Structure and Evidence: From Structured
Generation to Agentic Reasoning"
by
Mr. Zheye DENG
Abstract:
Large language models (LLMs) can produce fluent outputs that appear
trustworthy, but fluency cannot establish whether a structured artifact
respects its source or an agent's decision follows from its interaction
history. The failures take different forms. A plausible graph relation may
conflict with its surrounding graph, a well-formed table may omit source
information, and a successful decision may hide an unsupported sequence of
tool calls. This dissertation develops methods for judging such outputs and
actions through task structure and evidence that can be checked.
One line of work focuses on structured artifacts, covering the verification of
graph relations, the generation of tables from long documents, and the
evaluation of multiple visual formats. In commonsense knowledge graphs,
relation quality is assessed through edge semantics, mined logical rules, and
local neighborhoods, while an intermediate tuple representation for
long-document summarization exposes event selection, deduplication, and
aggregation before the final table is produced. Evaluation extends to tables,
graphs, and charts through separate criteria for faithfulness and coherence,
both validated against human ratings.
The second line studies agentic reasoning, where evidence accumulates through
interaction. A financial decision policy is trained with market outcomes and
process checks on its information requests and tool calls. A competitive
programming verifier returns a verdict and diagnosis before constructing an
executable counterexample for a suspected fault. If the counterexample is
legal and breaks the candidate under execution, it supplies concrete evidence
of the program failure. Across both lines, task-specific evidence anchors each
judgment and allows it to be checked independently of the model's own
explanation.
Date: Wednesday, 16 September 2026
Time: 9:00am - 11:00am
Venue: Room 3494
Lift 25/26
Committee Members: Dr. Yangqiu Song (Supervisor)
Prof. Raymond Wong (Chairperson)
Dr. Chaojian Li