Title: Structured Reasoning in LLMs

Presented by Malia Barker, Computing PhD student, Data Science emphasis

Abstract

Large language models (LLMs) exhibit strong linguistic competence but remain brittle on tasks requiring multi-step, compositional reasoning and reliable intermediate calculations. This work surveys inference-time approaches that enhance reasoning without modifying model parameters, tracing a progression from token-based prompting (which externalizes thoughts in natural language), to structured decomposition and search (which organize and explore alternatives), to programmatic and executable methods (which attach machine-checkable signals), and finally to agentic and reflective loops (which couple planning with tool use and feedback).

Across these families, we analyze trade-offs through three lenses—structure, verifiability, and agency—and synthesize empirical trends: making the process explicit consistently improves accuracy and interpretability, but most systems still validate linguistically and rely on ad hoc, format-agnostic traces that are difficult to certify, reuse, or standardize. This synthesis argues for a unifying perspective of reasoning as procedure: typed, modular representations whose steps are executable or validator-checkable, amenable to search and optimization at inference time, and portable across tasks without retraining. The review concludes with open problems and a research agenda spanning shared procedural schemas, process-level supervision, and benchmarks that evaluate the robustness of reasoning processes, not only final answers.

Advisor: Dr. Edoardo Serra

Committee Members: Dr. Francesco Gullo, Dr. Marion Scheepers

External Examiner: Dr. Grady Wright