PhD Dissertation Defense
University of South Carolina
Generalized Planning Using Language Models and Its Applications
Five research questions
Language models and planning
Neuro-symbolic AI
Dissertation Overview
Overview
A dissertation-scale summary of the problem formulation, architectural choices, and empirical findings.
The central claim of the dissertation is that language-model-based planning should be formulated as a constrained computational problem over explicit symbolic structure rather than as unconstrained autoregressive text generation.
Accordingly, the work combines taxonomy construction, controlled benchmarking, state-centric prediction, graph-based representations, symbolic successor validation, and metacognitive routing to identify where learned components are beneficial and where algorithmic structure is necessary.
Problem setting
Generalized planning is treated as reasoning over state transitions, validity constraints, and
transfer beyond the training distribution.
Architectural principle
Language models act as proposers, predictors, or representational components, while symbolic
modules provide verification, repair, search, and execution semantics.
Detailed structure
View the full research questions and
chapter-level contributions.
Technical TL;DR
Problem formulation
Standalone language-model generation lacks guarantees on plan validity, executability, and
systematic generalization across planning domains.
Systems thesis
More reliable performance emerges when learned components are embedded inside structured planning
pipelines with explicit states, constrained decoding, symbolic verification, and recovery
mechanisms.
Methodological contribution
The dissertation moves from descriptive analysis to constructive model and architecture design,
then evaluates those designs in both benchmark and application settings.
Overall conclusion
Hybrid neuro-symbolic planning architectures are more dependable than pure generation pipelines
when correctness, robustness, and domain transfer are first-class requirements.
Research Arc
1
Characterization
Construct a taxonomy of functional language-model roles in planning and analyze category drift across the literature.
2
Specialization
Evaluate valid plan generation and characterize the limitations of prompting and fine-tuning under distribution shift.
3
State-centric modeling
Replace action-sequence generation with successor-state prediction and symbolic successor selection.
4
Neuro-symbolic integration
Integrate learned proposers with symbolic verification, repair, and metacognitive control.
5
Applications
Evaluate the resulting methods in dialog systems, trustworthy assistance, and adaptive manufacturing.
Committee
Committee Members
Major Professor
Biplav Srivastava
University of South Carolina
Major Professor
Amit Sheth
University of South Carolina
Examination Chair
Ramtin Zand
University of South Carolina
Committee Member
Lior Horesh
IBM Research
Committee Member
Sarath Sreedharan
Colorado State University
