Abstract
LLM explanations are often fluent but unreliable, making them difficult to verify. This project converts LLM outputs into structured Claim–Evidence–Assumption graphs, improving verification accuracy and reducing false acceptance in debugging tasks.
Skills
Awards
Link
Contents
This is a 5 people course project! :) System & Technical Contributions
Built an LLM-to-graph pipeline using structured outputs and Pydantic to convert natural language explanations into Claim–Evidence–Assumption (CEA) graphs
Developed an interactive visualization system with Cytoscape.js, enabling graph navigation and traceable evidence inspection
Designed an evaluation framework comparing text vs. graph explanations, improving decision accuracy by 15.9% and reducing false acceptance rates
My Contributions
Implemented components of the LLM-to-graph pipeline, focusing on structured data processing and validation
Built parts of the interactive UI, improving usability and clarity of reasoning visualization
Contributed to dataset construction and system integration across LLM processing, graph modeling, and frontend visualization