Abstract

This talk examines how large language models (LLMs) used in software engineering can rely on hidden shortcuts, and how explainability can help detect and mitigate them. I first discuss empirical evidence from in-file vulnerability localization, where LLMs exhibit a “lost-in-the-end” effect: as files grow larger, vulnerabilities appearing later in the file become harder to detect, suggesting that models may rely on learned regularities about where bugs usually occur rather than fully reasoning over the code context. I then turn to cognitive biases in software-engineering decision support, presenting PROBE-SWE, a dynamic Prolog-based benchmark that generates paired biased and unbiased software-engineering dilemmas while controlling task logic and reasoning complexity. Building on this benchmark, I show why common prompting strategies offer limited protection, and how explicit software-engineering best-practice cues can reduce bias sensitivity. Finally, I discuss how global explainability can move from detecting failures to describing their triggers, introducing RuleSHAP as a method that combines SHAP-based global attribution with rule extraction to recover nonlinear behavioural patterns in LLMs.