ACL 2026 | AI润色还是代写？中科院四分类检测框架开源，锁定真实作者

TL;DR - RACE is an open-source, fine-grained detector that distinguishes human writing, LLM generation, AI-polished human text, and human-edited LLM text. It matters because these cases carry different implications for academic integrity but are conflated by existing binary and three-way detectors.

• RACE models rhetorical structure as a creator’s persistent “logical fingerprint” while treating surface language as the editor’s trace.
• It converts RST trees into multi-relational graphs and uses RoBERTa features, information-bottleneck projection, RGCNs, and supervised contrastive learning.
• On HART, RACE achieved 97.99% macro-AUROC and 83.06% average TPR at 1% FPR, outperforming the strongest reported baseline by 3.36 percentage points on the latter metric.
• Results remained strong across domains and RST parsers, though evaluation used only one dataset and performance is not yet sufficient for standalone commercial deployment.