TriagerX: Dual-Transformer Architecture Enhances Bug Triaging with Content and Interaction Analysis
TriagerX introduces a dual-transformer architecture to improve bug triaging tasks, addressing limitations in current pretrained language models (PLMs). Unlike single-transformer state-of-the-art baselines, TriagerX collects recommendations from two transformers, each using its last three layers to generate a robust content-based ranking. This approach better captures token semantics compared to traditional machine learning models that rely on statistical features like TF-IDF or bag of words. PLMs often attend to less relevant tokens in bug reports, reducing their effectiveness. Additionally, ignoring the interaction history of developers around similar bugs can lead to sub-optimal recommendations. The model, detailed in arXiv preprint 2508.16860v2, aims to enhance accuracy by considering both content and developer interactions. Its dual-transformer setup provides a more reliable assessment of token semantics, potentially improving bug assignment in software development.
Key facts
- TriagerX uses a dual-transformer architecture for bug triaging tasks
- It addresses limitations of pretrained language models (PLMs) that attend to less relevant tokens
- The model considers both content-based rankings and developer interaction history
- Unlike single-transformer baselines, it collects recommendations from two transformers
- Each transformer uses its last three layers for recommendations
- Traditional ML models rely on statistical features like TF-IDF and bag of words
- The research is documented in arXiv preprint 2508.16860v2
- The approach aims to improve bug report assignment accuracy
Entities
—