Extend evaluation to full fine-tuning and additional document genres

Extend the evaluation of conditioned cross-granularity training beyond LoRA fine-tuning and the studied receipt and scanned-business-form genres, including full fine-tuning and further document genres.

Background

The empirical study evaluates Qwen3-VL models at two scales using LoRA fine-tuning and three document corpora: two receipt datasets and one scanned business-form dataset. The authors explicitly identify full-parameter fine-tuning and additional document genres as unresolved extensions, which would test whether the reported reinforcement and collapse-avoidance effects generalize beyond the present training regime and domains.

References

Our evidence covers two receipt corpora and one corpus of scanned business forms, one VLM family at two scales and LoRA fine-tuning; full fine-tuning and further genres are left open, and the FUNSD corpus has $149$ training and $50$ test documents, so its intervals are correspondingly wide, with three seeds run only for mixed, conditioned and the neutral control.

— Joint Training Is Not Enough: Conditioned Cross-Granularity Training for Multimodal Document Understanding  (2609.00756 - Gan et al., 1 Sep 2026) in Limitations section