TVIR: Building Deep Research Agents Towards Text--Visual Interleaved Report Generation

2026-06-01

AI Summary

This work addresses the limitations of existing deep research agents, which are largely confined to generating text-only reports and lack effective evaluation of the factual accuracy of visual elements and their alignment with analytical content. To bridge this gap, we propose TVIR-Agent, the first generation and evaluation framework tailored for interleaved text-and-visual deep research reports. TVIR-Agent employs a hierarchical multi-agent architecture that collaboratively constructs outlines, retrieves relevant images, generates traceable charts, and performs context-aware interleaved writing. We also introduce TVIR-Bench, a benchmark featuring a dual-path joint text-and-visual evaluation mechanism that emphasizes the role of visual content in supporting specific analytical subgoals. Experiments across nine systems on TVIR-Bench demonstrate that TVIR-Agent achieves substantial performance gains, underscoring the critical importance of explicit multimodal design in evidence-driven report generation.

Abstract

Deep Research Agents have shown strong capability in multi-step information retrieval, reasoning, and long-form report generation, but existing benchmarks and systems remain predominantly text-centric, with limited evaluation of whether visual elements are factually reliable and well aligned with the surrounding analysis. To address this gap, we introduce TVIR (Text--Visual Interleaved Report Generation), which includes TVIR-Bench, a benchmark of 100 expert-curated multimodal deep research tasks that require visual elements to serve specific analytical sub-goals, and TVIR-Agent, a hierarchical multi-agent framework that serves as a strong baseline for constructing outlines, retrieving images, generating charts with traceable sources, and composing reports through context-aware sequential writing. We further develop a dual-path evaluation framework that combines Textual Assessment and Visual Assessment. Experiments across nine deep research systems show that TVIR-Agent achieves strong overall performance, underscoring the importance of explicit multimodal design and evaluation for evidence-driven report generation.

Problem

• Deep Research Agents
• Text-Visual Interleaved Report Generation
• Multimodal Evaluation
• Visual Reliability
• Evidence-Driven Report Generation

Innovation

• multimodal report generation
• deep research agents
• text–visual interleaving
• hierarchical multi-agent framework
• dual-path evaluation