Abstract

Before a bug report can be assigned to anyone, somebody has to decide what kind of problem it describes, how serious it is, how soon it needs attention and which component it touches. In busy projects this triage work piles up, and two people reading the same report often disagree. Large Language Models (LLMs) classify text well, but they are costly to run and awkward to deploy for everyday triage. Small Language Models (SLMs) need far less hardware, yet nobody has shown that they cope with the noisy, code-filled text of real bug reports. This paper reviews ten recent studies and proposes a design in which an SLM suggests labels and a short summary, a rule-based layer checks them, and reports the model is unsure about go to a human triager. We also set out how the design should be evaluated. Nothing has been implemented, so the paper makes no claim about accuracy or speed.

Keywords

How to Cite This Article

Rights