Two additional papers accepted at SBES'26!

Two additional papers were accepted at SBES 2026!

Evaluator Bias in LLM-Based Program Repair: A Disagreement Analysis on Android Resource Leaks reports the current research conducted by Andrezza Bonfim (CIn-UFPE) for her master’s dissertation. In this work, we investigate how the choice of evaluation method affects conclusions about the effectiveness of LLM-generated program repairs. Using Android resource leaks as our study context, we compare GPT-5.4, Claude 4.6, and Gemini 3.1 across 300 repair scenarios and evaluate their outputs using lexical, structural, and neural semantic techniques. The results show substantial disagreement among evaluation methods, reinforcing that simple textual similarity can underestimate valid repairs and that evaluating AI-generated code requires techniques that better account for program structure and semantics.

Moreover, Dependency Version Migration in Software Projects: A Rapid Review of Motivations, Techniques, Tools, and Challenges is a collaboration with Arthur Vinícius Bezerra da Silva (IFPE), Bruno Cartaxo (IFPE & Zup Innovation), and Wesley Assunção (NCSU). We reviewed 63 primary studies to characterize why software projects migrate dependency versions, how migration needs are identified, which techniques and tools support the process, and what challenges remain. Our findings reveal a fragmented research scenario: migrations are commonly driven by security, bug fixes, and compatibility concerns, while existing automated approaches remain highly ecosystem-specific and face recurring trade-offs between precision, generalizability, and manual effort.

See you at CBSoft 2026!