ReFeedGraph: Graph-grounded Reusable Feedback for Debate Improvement

Research Problem

ReFeedGraph is a graph-grounded LLM feedback system that converts debate critiques into reusable revision principles for improving weaker debate arguments.

Research Contribution

• Integrates feedback from multiple debate-judging frameworks.
• Stores failure–repair patterns in a reusable graph.
• Retrieves and adapts relevant revision principles.
• Generates targeted feedback for debate participants.
• Evaluated on DebateArt, IQ2, and British Parliamentary datasets.

Methodology and Evaluation

The system extracts generalizable principles from judge feedback, organizes them in a graph, and retrieves relevant principles to generate targeted revisions for future debates. ReFeedGraph was evaluated on the DebateArt, IQ2, and British Parliamentary datasets. It achieved:

• Up to 0.491 average score improvement over the original weaker-side arguments.
• Up to 85.1% improved cases across the evaluated datasets.
• Up to 89.7% winner-flip rate, indicating that the revised argument was judged stronger than the original winning side in a substantial portion of cases.

Technologies

Large language models, knowledge graphs, retrieval-augmented generation, debate analysis, natural language processing, Python.