Safety
DocSeeker: Structured Visual Reasoning with Evidence Grounding for Long Document Understanding
arXiv:2604.12812v1 Announce Type: new Abstract: Existing Multimodal Large Language Models (MLLMs) suffer from significant performance degradation on the long document understanding task as document le
arXiv:2604.12812v1 Announce Type: new Abstract: Existing Multimodal Large Language Models (MLLMs) suffer from significant performance degradation on the long document understanding task as document length increases. This stems from two fundamental challenges: 1) a low Signal-to-Noise Ratio (SNR), with crucial evidence buried in irrelevant pages; and 2) supervision scarcity, as datasets offering only final short answers provide a weak learning signal. In this paper, we address these challenges by proposing a paradigm that requires the model to execute a structured ``extbf{Analysis}, extbf{Localization} and extbf{Reasoning}'' workflow. To instill this capability, we design a two-stage training framework: we first perform Supervised Fine-Tuning on high-quality data generated via an efficient knowledge distillation strategy. Subsequently, we employ an Evidence-aware Group Relative Policy Optimization which jointly optimizes for both evidence localization and answer accuracy. Additionally, we introduce a Evidence-Guided Resolution Allocation strategy to mitigate memory constraints of training on multi-pages documents. Extensive experiments demonstrate that DocSeeker achieves superior performance on both in-domain and out-of-domain tasks. We show it robustly generalizes from short-page training to ultra-long documents and is naturally synergistic with visual Retrieval-Augmented Generation systems, serving as a solid foundation for their implementation.
Related
- VideoStir: Understanding Long Videos via Spatio-Temporally Structured and Intent-Aware RAG
- SVD-Prune: Training-Free Token Pruning For Efficient Vision-Language Models
- MODIX: A Training-Free Multimodal Information-Driven Positional Index Scaling for Vision-Language Models
- VISOR: Agentic Visual Retrieval-Augmented Generation via Iterative Search and Over-horizon Reasoning
- Reasoning Within the Mind: Dynamic Multimodal Interleaving in Latent Space
Source: arXiv cs.AI | 2026-04-15