Engineering · 4 min read · April 20, 2026
Dual Transformers Improve Bug Assignment Accuracy by 10%+
TriagerX uses two transformer models and developer interaction history to recommend the right engineer for bug fixes, outperforming single-model approaches.
TriagerX pairs two transformers with interaction-based ranking to assign bugs to developers more accurately than single-model baselines.
- — Dual-transformer design extracts recommendations from six layers total, reducing noise from irrelevant tokens.
- — Interaction-based ranking layer refines assignments by examining developer history with similar resolved bugs.
- — Achieves 10%+ improvement in Top-1 and Top-3 accuracy across five public datasets.
- — Deployed in production at large software company for both developer and component assignment.
- — Outperforms nine existing transformer-based methods including published state-of-the-art systems.
- — Component recommendations improved 10%; developer recommendations improved 54% in production setting.
- — Addresses limitation of PLMs attending to irrelevant tokens in unstructured bug reports.
Frequently asked
- TriagerX uses two transformers instead of one, each contributing recommendations from its last three layers (six total). This ensemble approach reduces the impact of attending to irrelevant tokens in bug reports. The dual design captures complementary semantic views, improving robustness. Single-transformer baselines rely on one model's attention pattern, which can miss nuance in unstructured bug text.