“Bring the image back!” An Empirical Study of Visual Summaries in Code Review

Luís F. Gomes, Xin Zhou, Vincent Hellendoorn, Jonathan Aldrich, Rui Abreu, David Lo

Preprint — under review 2026

JupyterDraw automatically generates informal diagrams from Jupyter notebooks to help reviewers catch silent flaws (data leakage, improper evaluation) in AI-generated code. A controlled experiment with 26 participants found the visual overview caught +0.50 more flaws per task than a textual one, raising the share of reviews catching both planted flaws from 29% to 67%.

SE-Jury: An LLM-as-Ensemble-Judge Metric for Narrowing the Gap with Human Evaluation in SE

Xin Zhou, Kisub Kim, Ting Zhang, Martin Weyssow, Luís F. Gomes, Guang Yang, Kui Liu, Xin Xia, David Lo

CORE A* International Conference on Automated Software Engineering (ASE) 2025

SE-Jury is an LLM-as-ensemble-judge metric that combines five distinct evaluation strategies to assess the functional correctness of generated software artifacts, with a dynamic team selection mechanism that picks the optimal subset of evaluators — cutting LLM API costs by ~50% while maintaining evaluation accuracy.

[arXiv]

VisDocSketcher: Towards Scalable Visual Documentation with Agentic Systems

Luís F. Gomes, Xin Zhou, David Lo, Rui Abreu

arXiv preprint 2025

An agentic LLM system that generates high-level visual documentation from source code, paired with AutoSketchEval — a reference-free evaluation framework (inspired by autoencoder reconstruction) that scores diagram quality with no ground truth, reaching AUC > 0.87 across 1,000 Jupyter notebooks.

[arXiv]