Crag Wolfe

Renyu Li, Vladimir Kirilenko, Yao You et al.

Fine-tuning object detection (OD) models on combined datasets assumes annotation compatibility, yet datasets often encode conflicting spatial definitions for semantically equivalent categories. We propose an agentic label harmonization workflow that uses a vision-language model to reconcile both category semantics and bounding box granularity across heterogeneous sources before training. We evaluate on document layout detection as a challenging case study, where annotation standards vary widely across corpora. Without harmonization, naïve mixed-dataset fine-tuning degrades a pretrained RT-DETRv2 detector: on SCORE-Bench, which measures how accurately the full document conversion pipeline reproduces ground-truth structure, table TEDS drops from 0.800 to 0.750. Applied to two corpora whose 16 and 10 category taxonomies share only 8 direct correspondences, harmonization yields consistent gains across content fidelity, table structure, and spatial consistency: detection F-score improves from 0.860 to 0.883, table TEDS improves to 0.814, and mean bounding box overlap drops from 0.043 to 0.016. Representation analysis further shows that harmonized training produces more compact and separable post-decoder embeddings, confirming that annotation inconsistency distorts the learned feature space and that resolving it before training restores representation structure.

Renyu Li, Antonio Jimeno Yepes, Yao You et al.

Multi-modal generative document parsing systems challenge traditional evaluation: unlike deterministic OCR or layout models, they often produce semantically correct yet structurally divergent outputs. Conventional metrics-CER, WER, IoU, or TEDS-misclassify such diversity as error, penalizing valid interpretations and obscuring system behavior. We introduce SCORE (Structural and COntent Robust Evaluation), an interpretation-agnostic framework that integrates (i) adjusted edit distance for robust content fidelity, (ii) token-level diagnostics to distinguish hallucinations from omissions, (iii) table evaluation with spatial tolerance and semantic alignment, and (iv) hierarchy-aware consistency checks. Together, these dimensions enable evaluation that embraces representational diversity while enforcing semantic rigor. Across 1,114 pages spanning a holistic benchmark and a field dataset, SCORE consistently revealed cross-dataset performance patterns missed by standard metrics. In 2-5% of pages with ambiguous table structures, traditional metrics penalized systems by 12-25% on average, leading to distorted rankings. SCORE corrected these cases, recovering equivalence between alternative but valid interpretations. Moreover, by normalizing generative outputs into a format-agnostic representation, SCORE reproduces traditional scores (e.g., table F1 up to 0.93) without requiring object-detection pipelines, demonstrating that generative parsing alone suffices for comprehensive evaluation. By exposing how interpretive diversity impacts evaluation outcomes and providing multi-dimensional, interpretable diagnostics, SCORE establishes foundational principles for semantically grounded, fair, and practical benchmarking of modern document parsing systems.