mlsanityData to 14 Oct 2026

AutothinkRAG: Complexity-Aware Control of Retrieval-Augmented Reasoning for Image-Text Interaction

arXiv 2603.05551 ↗First posted 5 Mar 2026cs.IR cs.CVcited by 0

On the map: LLM reasoning → Multimodal reasoning

also in 1 more
Abstract (arXiv)

Multimodal document question answering requires retrieving dispersed evidence from visually rich long documents and performing reliable reasoning over heterogeneous information. Existing multimodal RAG systems remain limited by two bottlenecks: static retrieval that ignores query complexity, and end-to-end Vision-Language Models (VLMs) that couple visual perception with logical reasoning, leading to inefficient computation and unstable answer generation. We propose AutoThinkRAG, a complexity-aware inference architecture for multimodal document QA. It has two components: (1) a Query Complexity Router that analyzes query difficulty and structure to adaptively select retrieval and reasoning paths; and (2) a Perception--Reasoning Decoupling architecture that uses a lightweight VLM as a high-fidelity visual interpreter to convert query-relevant visual cues into textual representations, which are then passed to an LLM for logical reasoning and answer synthesis. This design improves both efficiency and robustness, especially on long-document and unanswerable queries. Experiments on DocBench and MMLongBench show that AutoThinkRAG achieves 82.13% and 51.29% overall accuracy, respectively, while reducing per-query token consumption by 18.9% and monetary cost by 18.2%. Further analyses show that the gains are most pronounced on complex queries requiring adaptive retrieval and multi-step reasoning.

Key facts. No citing papers in the map yet (data through 14 Oct 2026). Placed in Multimodal reasoning, one of 17 directions in LLM reasoning. Builds on 28 of its 45 references that are in the map. First posted on arXiv 5 Mar 2026.

Builds on

28 of 45 references.

References · 28

  1. 2024From Local to Global: A Graph RAG Approach to Query-Focused Summarizationcited by 1,453
  2. 2023Enhancing Retrieval-Augmented Large Language Models with Iterative Retrieval-Generation Synergycited by 346
  3. 2024LightRAG: Simple and Fast Retrieval-Augmented Generationcited by 315
  4. 2024VisRAG: Vision-based Retrieval-augmented Generation on Multi-modality Documentscited by 245
  5. 2022LinkBERT: Pretraining Language Models with Document Linkscited by 211
  6. 2024ColPali: Efficient Document Retrieval with Vision Language Modelscited by 195
  7. 2024GNN-RAG: Graph Neural Retrieval for Large Language Model Reasoningcited by 128
  8. 2022REVEAL: Retrieval-Augmented Visual-Language Pre-Training with Multi-Source Multimodal Knowledge Memorycited by 123
  9. 2024MMed-RAG: Versatile Multimodal RAG System for Medical Vision Language Modelscited by 85
  10. 2025VideoRAG: Retrieval-Augmented Generation with Extreme Long-Context Videoscited by 77
  11. 2025HyperGraphRAG: Retrieval-Augmented Generation via Hypergraph-Structured Knowledge Representationcited by 68
  12. 2025PathRAG: Pruning Graph-based Retrieval Augmented Generation with Relational Pathscited by 61
16 more
  1. 2016Dynamic Coattention Networks For Question Answeringcited by 49
  2. 2023Retrieval-Enhanced Contrastive Vision-Text Modelscited by 36
  3. 2025ArchRAG: Attributed Community-based Hierarchical Retrieval-Augmented Generationcited by 28
  4. 2024RAGraph: A General Retrieval-Augmented Graph Learning Frameworkcited by 24
  5. 2025HopRAG: Multi-Hop Reasoning for Logic-Aware Retrieval-Augmented Generationcited by 24
  6. 2025RAG-Anything: All-in-One RAG Frameworkcited by 20
  7. 2025DocAgent: A Multi-Agent System for Automated Code Documentation Generationcited by 18
  8. 2024Document Haystacks: Vision-Language Reasoning Over Piles of 1000+ Documentscited by 17
  9. 2024TRACE the Evidence: Constructing Knowledge-Grounded Reasoning Chains for Retrieval-Augmented Generationcited by 16
  10. 2025MMGraphRAG: Bridging Vision and Language with Interpretable Multimodal Knowledge Graphscited by 16
  11. 2025BookRAG: A Hierarchical Structure-aware Index-based Approach for Retrieval-Augmented Generation on Complex Documentscited by 16
  12. 2025DocLens : A Tool-Augmented Multi-Agent Framework for Long Visual Document Understandingcited by 15
  13. 2024GarmentAligner: Text-to-Garment Generation via Retrieval-augmented Multi-level Correctionscited by 10
  14. 2022Capturing Global Structural Information in Long Document Question Answering with Compressive Graph Selector Networkcited by 9
  15. 2025Beyond Chunking: Discourse-Aware Hierarchical Retrieval for Long Document Question Answeringcited by 7
  16. 2023MemSum-DQA: Adapting An Efficient Long Document Extractive Summarizer for Document Question Answeringcited by 1

Built on by

Not cited yet.

Similar, not cited

Close in wording, no citation either way

  1. 2024VisDoM: Multi-Document QA with Visually Rich Elements Using Multimodal Retrieval-Augmented Generation
  2. 2026Stop Thinking, Start Looking: Efficient Post-Training for Multimodal Document Question Answering via Reasoning-Free Alignment
  3. 2025Progressive Multimodal Search and Reasoning for Knowledge-Intensive Visual Question Answering
  4. 2026Doc-REFRAG: Rethinking Multimodal Document Retrieval-Augmented Generation
  5. 2025ReAG: Reasoning-Augmented Generation for Knowledge-based Visual Question Answering
  6. 2025Benchmarking Retrieval-Augmented Multimodal Generation for Document Question Answering
  7. 2026TAP-RAG: Task-Aware Policy Control for Long-Document Multimodal Question Answering
  8. 2025QA-Dragon: Query-Aware Dynamic RAG System for Knowledge-Intensive Visual Question Answering