RTE4SDC at the ICST 2026 Tool Competition—Self-Driving Car Testing Track

Test case prioritisation for self-driving cars (SDCs) aims to reorder test suites so that failure-revealing tests run as early as possible. Existing approaches target test selection, and while they can be adapted for prioritisation by sorting tests by predicted failure probability, a classifier optimised for per-item accuracy does not necessarily produce a good ranking. This paper presents RTE4SDC, a Transformer Encoder trained with a hybrid classification and pairwise ranking loss, designed to optimise test ordering directly while retaining the ability to distinguish failing from passing tests. Evaluated on the heldout 10% test partition of SensoDat, RTE4SDC achieves a mean APFD of 0.76 over 10 repeated runs, compared to 0.79 for the competition baseline (ITEP4SDC). This gap is likely influenced by the evaluation setting, as RTE4SDC is assessed under a 70/20/10 train/validation/test split, whereas the competition submission uses an 80/20 train/validation split on SensoDat without a heldout test set, which is expected to narrow the difference.

ICST 2026 Tool Competition