DocHRL: Adaptive Document Classification via Hierarchical Reinforcement Learning
A new framework called DocHRL uses hierarchical reinforcement learning to optimize document classification by dynamically selecting cost-effective policies per document. It addresses inefficiency in traditional pipelines that apply uniform model sequences regardless of document complexity. DocHRL formulates classification as a sequential decision problem with a two-level hierarchy: a top-level policy chooses among broad options like vision classifiers, LLMs, OCR, or human review, while sub-policies select specific models. The reward signal minimizes total expected cost including inference, misclassification, and human labeling costs. The framework is trained with PPO and aims to reduce resource waste in real-world document processing.
Key facts
- DocHRL is a hierarchical reinforcement learning framework for document classification.
- It adaptively selects cost-effective policies per document.
- Two-level policy hierarchy: top-level selects broad options, sub-policies choose concrete models.
- Reward signal is negative total expected cost including inference, misclassification, and human labeling.
- Trained with PPO.
- Addresses inefficiency of uniform model sequences in traditional pipelines.
- Options include vision classifiers, LLMs, OCR, and human-in-the-loop review.
- Published on arXiv with ID 2607.22644.
Entities
Institutions
- arXiv