From Conceptualization to Metaphysical Reasoning: Frameworks and Benchmarks Towards Generalizable Reasoning in Large Language Models

PhD Thesis Proposal Defence


Title: "From Conceptualization to Metaphysical Reasoning: Frameworks and
Benchmarks Towards Generalizable Reasoning in Large Language Models"

by

Mr. Weiqi WANG


Abstract:

Large language models (LLMs) show strong reasoning across many tasks, yet
their reliability can vary when assumptions change, inputs shift distribution,
or familiar knowledge must be recomposed in novel ways. This thesis argues
that a key organizing principle for improving such generalization is
conceptualization: the ability to abstract concrete events and entities into
reusable concepts, and to instantiate those concepts in new situations. We
develop a unified paradigm in which conceptualization structures how
commonsense knowledge is represented, acquired, scaled, modified, and
evaluated beyond surface competence.

We first systematize concept-centric methods for LLMs and formalize a
conceptualization–instantiation cycle over commonsense knowledge bases (CSKBs)
as a lens spanning generation, question answering, and knowledge
manipulation. Building on this lens, we introduce approaches that construct
and exploit concept-structured event and entity knowledge to improve
generative commonsense inference and zero-shot commonsense question
answering, showing that concept-level structure can strengthen reasoning
without relying solely on model scale. To address limited CSKB coverage, we
propose a scalable distillation framework that extracts large volumes of
concept-structured knowledge from strong LLMs and uses critic-style filtering
to retain plausible, useful knowledge, expanding coverage while preserving
quality.

Beyond acquisition, we study controlled knowledge modification via a
concept-level editing framework that couples automated plausibility
verification with concept-aware rewriting, improving both factuality and
downstream utility. Finally, we introduce metaphysical reasoning as a
concept-driven stress test: reasoning about improbable or counterfactual
changes to conceptualized events. We provide a benchmark that decomposes this
challenge into discriminating event, inference, and transition validity under
controlled distribution shifts, revealing persistent gaps between apparent
competence and deeper conceptual understanding. Together, these contributions
advance frameworks, resources, and evaluations that push LLMs toward more
robust, generalizable reasoning grounded in concept-level structure.


Date:                   Thursday, 5 February 2026

Time:                   12:00noon - 2:00pm

Venue:                  Room 3494
                        Lift 25/26

Committee Members:      Dr. Yangqiu SONG (Supervisor)
                        Prof. Raymond WONG (Chairperson)
                        Dr. May FUNG