Multi-view fine-tuning of UniXcoder yields 0.845 macro F1 for binary machine-generated code detection on unseen languages and domains, while class-weighted training is required to lift multi-class attribution macro F1 from 0.086 to 0.345 under 221:1 imbalance.
Yifan Wang, Yujie Li, Rui Zhang, and Minghao Chen
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.SE 1years
2026 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
UCSC-NLP at SemEval-2026 Task 13: Multi-View Generalization and Diagnostic Analysis of Machine-Generated Code Detection
Multi-view fine-tuning of UniXcoder yields 0.845 macro F1 for binary machine-generated code detection on unseen languages and domains, while class-weighted training is required to lift multi-class attribution macro F1 from 0.086 to 0.345 under 221:1 imbalance.