TVIR: Breaking Text‑Only Limits with AI‑Powered Multimodal Research Report Generation

Background

Large language models enable Deep Research Agents that perform multi‑step retrieval, reasoning, and generation to produce full research reports. Existing benchmarks evaluate only textual coherence, analytical depth, and citation support, ignoring visual evidence that professional reports require.

TVIR‑Bench

TVIR‑Bench is a multimodal research‑report benchmark containing 100 expert‑curated tasks (50 Chinese, 50 English) across 10 domains and 3 difficulty levels . Task design follows five principles—role‑driven, demand‑oriented, deep research, frontier focus, multimodal integration—to ensure real‑world relevance and substantive analysis.

Task construction uses a four‑stage expert workflow:

1. Expert topic proposal: domain experts suggest timely, novel topics.

2. LLM‑based drafting: a large language model expands the topic into a coherent draft.

3. Multi‑expert review: three experts check design compliance, factual accuracy, logical flow, and multimodal effectiveness.

4. Checklist compilation: an evaluation checklist records required components for systematic verification.

TVIR‑Agent Framework

3.1 Research‑driven Planning

The Planner parses the user task, iteratively calls search and web‑scraping tools, and synthesises retrieved information into a structured outline. Each outline unit records:

Chapter title and abstract

Planned visual requirements

Research notes (citations, source URLs, key findings)

3.2 Visual Resource Instantiation

Two specialised agents fulfil visual needs:

Image Searcher: retrieves candidate images (portraits, diagrams, architecture sketches) via Google Image Search and validates relevance with a VQA tool.

Chart Generator: gathers data for distribution‑ or relationship‑focused content, checks cross‑source consistency, generates Python plotting code, and executes it in a sandbox.

3.3 Context‑aware Sequential Writing

The Writer generates the report chapter by chapter. It conditions on the current outline unit and a dynamically updated global context (titles, abstracts, sub‑section structures of already written chapters) to maintain cross‑chapter coherence and avoid redundancy. The Writer consumes chapter‑level research notes as evidence, invokes search tools when evidence is insufficient, and inserts visual assets at appropriate points to produce interleaved Markdown.

3.4 Global Index Polishing

The Polisher operates at the report level. It removes unreferenced bibliography entries, deduplicates URLs and content, renumbers references into a unified list, and updates in‑text citation markers. It also renumbers all images, assigns sequential IDs and captions, and updates image references in the body.

Dual‑Path Evaluation Framework

Evaluation consists of two complementary components:

Textual Assessment (TA): measures coherence, analytical depth, and citation support.

Visual Assessment (VA): measures relevance, correctness, and integration of visual evidence.

Experimental Results

Nine deep‑research systems were evaluated: six commercial systems and three TVIR‑Agent variants.

Overall performance: TVIR‑Agent (Claude‑4.5‑Sonnet) achieved the highest overall score, followed by TVIR‑Agent (Qwen3‑Max) and TVIR‑Agent (GLM‑4.7). Among commercial systems, Manus‑1.6 ranked highest.

Variant strengths: GLM‑4.7 obtained the highest TA score, demonstrating strong textual ability; Claude‑4.5‑Sonnet obtained the highest VA score, showing superior visual support and cross‑modal alignment.

Citation support gap: GLM‑4.7 scored 68.64 on the Citation Support metric, 21.11 points above the best commercial system (Claude‑4.5‑Sonnet w/Search, 47.53).

Task complexity effect: As difficulty increased, Instruction Alignment scores generally declined, while Analytical Depth & Breadth scores rose, indicating higher multimodal coordination and long‑range reasoning demands.

Structural Error Analysis

Three error categories were examined:

Traceability errors: missing or inaccessible sources for facts, data, or figures.

Consistency errors: duplicate or missing numbering in internal indexes.

Completeness errors: unavailable or incomplete cited components such as broken charts or missing captions.

TVIR‑Agent variants produced significantly fewer structural errors than commercial systems, indicating better end‑to‑end control over citation management and multimodal assembly.

Tool Usage Analysis

Different TVIR‑Agent variants exhibited distinct tool‑allocation patterns:

GLM‑4.7: most frequent use of search and crawl tools, highest Average Effective Citations (AEC) of 102.41, but low chart completion (38.45 %). Average planned charts per task were 8.66, actual generated charts 3.33.

Claude‑4.5‑Sonnet: balanced strategy, maintaining AEC of 86.14 while achieving the highest chart completion rate of 94.61 %.

These results show that performance depends not only on the backbone model but also on how tools are allocated between retrieval and chart generation.

Conclusion

TVIR introduces a unified benchmark (TVIR‑Bench), a hierarchical multi‑agent framework (TVIR‑Agent), and a dual‑path evaluation framework for text‑visual interleaved report generation. Experiments demonstrate that TVIR‑Agent delivers strong overall performance, especially in evidence support and cross‑modal alignment, while current deep‑research systems remain weaker in visual integration.

Code example

本文
约2400字
，建议阅读
5
分钟
本文介绍 TVIR 体系，实现 AI 深度研究图文协同生成与双维度评测。