Broad four-way zero-shot generalization of mixed-origin detection

Determine whether D2C-Routing can achieve broad zero-shot generalization to four-way HH/HA/AH/AA mixed-origin detection across datasets and construction protocols, beyond the evaluated MixD2C setting.

Background

The paper evaluates transfer from the HART-derived MixD2C setting to external datasets and reconstructed benchmarks. Although targeted dimension-level transfer—especially expression-origin transfer—shows useful results, direct four-way transfer to the HART-style MixSet reconstruction is weak, with a Macro-F1 of 0.2262 and an AUROC of 0.4372.

Because the external datasets use different label spaces, construction procedures, and evaluation protocols, the reported experiments do not establish whether the D2C-Routing approach generalizes broadly without target adaptation. The authors therefore leave broad four-way zero-shot generalization unresolved.

References

Direct MixSet transfer remains weak, leaving broad four-way zero-shot generalization unresolved.

— D2C-Routing: Dimension-to-Composition Evidence Routing for Mixed-Origin AI-Generated Text Detection  (2608.27380 - Chen et al., 27 Aug 2026) in Appendix, Section "Transfer Checks", subsection "Transfer Checks"