FocalOrder: Focal Preference Optimization for Reading Order Detection

TL;DR

FocalOrder tackles reading order detection by addressing Positional Disparity—the mismatch between uniform supervision and the non-uniform layout difficulty—through Focal Preference Optimization. It introduces Adaptive Difficulty Discovery, which uses EMA to emphasize structurally ambiguous transitions, and Difficulty-Calibrated Pairwise Ranking, which imposes global coherence via adaptive margins in a ranking objective. The method achieves state-of-the-art results on OmniDocBench and Comp-HRDoc with a compact backbone, demonstrating that focusing optimization on hard regions improves robustness beyond larger models. This approach offers a scalable paradigm for document understanding by aligning training signals with intrinsic layout entropy, improving reliability in complex layouts and downstream tasks such as RAG pipelines.

Abstract

Reading order detection is the foundation of document understanding. Most existing methods rely on uniform supervision, implicitly assuming a constant difficulty distribution across layout regions. In this work, we challenge this assumption by revealing a critical flaw: \textbf{Positional Disparity}, a phenomenon where models demonstrate mastery over the deterministic start and end regions but suffer a performance collapse in the complex intermediate sections. This degradation arises because standard training allows the massive volume of easy patterns to drown out the learning signals from difficult layouts. To address this, we propose \textbf{FocalOrder}, a framework driven by \textbf{Focal Preference Optimization (FPO)}. Specifically, FocalOrder employs adaptive difficulty discovery with exponential moving average mechanism to dynamically pinpoint hard-to-learn transitions, while introducing a difficulty-calibrated pairwise ranking objective to enforce global logical consistency. Extensive experiments demonstrate that FocalOrder establishes new state-of-the-art results on OmniDocBench v1.0 and Comp-HRDoc. Our compact model not only outperforms competitive specialized baselines but also significantly surpasses large-scale general VLMs. These results demonstrate that aligning the optimization with intrinsic structural ambiguity of documents is critical for mastering complex document structures.

Figures (10)

• Figure 1: Illustration of Positional Disparity. While representative models demonstrate mastery over deterministic regions (start/end), they suffer from significant performance degradation in the document body. This reveals a misalignment between the uniform supervision used in training and the non-uniform structural complexity of real-world documents.
• Figure 2: Error rates of representative models across normalized document positions. The consistent "Inverted-U" curve across datasets (solid lines: OmniDocBench, dashed lines: Comp-HRDoc) reveals a systematic bias, i.e., models struggle to serialize the complex document body compared to the rigid start and end templates.
• Figure 3: Spatial-Logical Mismatch Analysis. Distribution of spatial-logical mismatches across relative positions on OmniDocBench v1.0 and Comp-HRDoc.
• Figure 4: Overview of the FocalOrder framework. The architecture integrates two components: Adaptive Difficulty Discovery, which leverages an EMA-based tracker to dynamically identify and up-weight ($w_t$) structurally ambiguous transitions; and Difficulty-Calibrated Pairwise Ranking, which implements contrastive optimization using a difficulty-aware advantage function and adaptive margins to prioritize complex layout patterns over trivial ones.
• Figure 5: Comparison of error distributions on OmniDocBench v1.0. Unlike the baseline, which suffers from the "Inverted-U" degradation, FocalOrder (green line) effectively flattens the curve, maintaining robust performance even in the complex intermediate sections.