School of Information Sciences

Meng Li's Preliminary Examination

Meng Li

PhD candidate Meng Li will present her dissertation proposal, "Toward Reliable, Explainable, and Generalizable Neurosymbolic AI Systems with Large Language Models." Li's preliminary examination committee includes Bertram Ludäscher, Timothy M. McPhillips, Shawn Bowers, Halil Kilicoglu, and Ismini Lourentzou.

Abstract

With the rise of Large language models (LLMs), AI has been widely used in diverse domains, and is increasingly relied on not only to provide answers but to support decision making. Acting on such answers requires deciding whether to trust them, which depends on both their reliability (whether the answers can be expected to be correct) and their explainability (whether the explanations of how they were derived are auditable and faithful). These properties are especially critical in high-stakes domains such as medicine and law. However, LLMs struggle with both: they often produce unreliable answers, and asking them to explain their answers may yield only plausible text rather than a faithful account of their actual reasoning. This thesis uses the notebook understanding task as a testbed for this kind of evidence-based judgment. The notebook understanding task is to recover how information flows through a computational notebook without executing it. This reduces to a sequence of decisions: for a given variable and a given cell, whether the variable is an input to the cell, and whether it is an output candidate of the cell. Each such decision has a well-defined ground truth, making notebook understanding a decision-support task that can be evaluated rigorously. This thesis develops a sequence of neurosymbolic AI systems for this task, each refining how an LLM is combined with explicit symbolic structure. CRABS constrains both the workload given to an LLM and the responses it can generate, which makes the system more reliable. ARIAS then externalizes the underlying reasoning into an argumentation framework (AF) by instantiating argument templates using facts from a knowledge base (KB), so that positive decisions come with auditable and faithful explanations. The proposed work contributes a reusable framework that automatically learns task-specific specifications for building KBs and libraries of argument templates, replacing the per-task human-designed ones required by ARIAS.

Questions? Contact Meng Li.

School of Information Sciences

501 E. Daniel St.

MC-493

Champaign, IL

61820-6211

Voice: (217) 333-3280

Email: ischool@illinois.edu

Back to top