Hierarchical multi-agent reinforcement learning for retrieval-augmented industrial document question answering.

Area of Science:

• Artificial Intelligence
• Information Retrieval
• Document Analysis

Background:

• Multimodal industrial documents integrate text, images, and layout for domain knowledge.
• Existing retrieval-augmented generation (RAG) frameworks use static policies, limiting adaptability to complex queries and cross-modal dependencies.
• This leads to incomplete evidence retrieval and suboptimal reasoning in long-document scenarios.

Purpose of the Study:

• To develop an adaptive framework for multimodal retrieval-augmented reasoning.
• To address the limitations of static retrieval and fusion policies in RAG frameworks.
• To improve reasoning performance on complex, long multimodal documents.

Main Methods:

• Proposed MARL-RAGDoc, a hierarchical multi-agent reinforcement learning framework.
• Implemented a coordinator agent for dynamic modality weighting and retrieval depth allocation.
• Utilized specialized agents for text, image, and table evidence selection and a collaborative reasoning module for policy optimization.

Main Results:

• MARL-RAGDoc demonstrated superior performance over baseline methods on multimodal document benchmarks.
• Achieved consistent improvements in both retrieval accuracy and reasoning performance.
• The framework proved to be computationally efficient.

Conclusions:

• MARL-RAGDoc effectively addresses the challenges of multimodal document understanding and reasoning.
• The adaptive, multi-agent approach enhances information retrieval and reasoning accuracy.
• The framework offers a computationally efficient solution for complex long-document scenarios.