Continuous Integration (CI) builds execute large test suites that often yield multiple simultaneous failures, many originating from a common underlying fault. Test failure triage -- grouping failing tests by their root cause -- is therefore critical for reducing debugging effort. Existing approaches work poorly in CI builds with multiple faults, and either lack semantic reasoning or ignore highly informative structural and change-aware signals. To address these limitations, we propose HyTri, a framework that integrates LLM-guided causal reasoning into test failure triage. HyTri derives similarity signals that consider failure semantics and hypothesized root causes. It fuses these signals using confidence weighting and error-type adjustment into a unified similarity measure that is used to cluster the failing tests. To evaluate HyTri we created Multi-Fault CI Build Benchmark (MFCI-Benchmark), a novel benchmark for test failure triage algorithms comprising synthesized multi-fault CI builds from well-known bug repositories. Experimental results over MFCI-Benchmark and the multifault Defects4J fault localization benchmark showed that HyTri outperforms existing baselines, maintaining strong performance even in datasets where stack traces are shallow and uninformative.

@inproceedings{paperbot12550,
  title = {HyTri: Hybrid Triage of CI Failures via LLM Guided Semantic Reasoning and Change Attribution},
  author = {Lior Broide and Argaman Mordoch and Roni Stern},
  booktitle = {ASE},
  year = {2026},
  doi = {10.1145/3832783.3837467}
}