TVIR: Text‑Visual Interleaved Report Generation Empowers Deep Research Agents

Background and Motivation

Rapid advances in large language models have led to the emergence of Deep Research Agents that can autonomously perform multi‑step information retrieval, reasoning, and generation to produce complete research reports. Existing evaluation benchmarks and systems are text‑centric, assessing coherence, analytical depth, and citation support while ignoring the crucial role of visual evidence in professional reports.

In practice, high‑quality research reports intertwine narrative analysis with figures, charts, and images that serve as core evidence for arguments, trends, and comprehension. Even systems that include visual elements often treat them as decorative rather than integral to reasoning.

TVIR: Benchmark and Framework

TVIR (Text‑Visual Interleaved Report Generation) is a unified benchmark and agent framework proposed by Nanjing University and Alibaba. It systematically evaluates multimodal capabilities of deep research agents.

TVIR‑Bench

TVIR‑Bench is the first comprehensive benchmark designed for end‑to‑end multimodal report generation. It contains 100 expert‑curated tasks (50 Chinese, 50 English) covering 10 domains and three difficulty levels. Task design follows five core principles: role‑driven, demand‑oriented, deep‑research, frontier‑focused, and multimodal integration, ensuring alignment with real user needs and requiring substantive analysis rather than simple retrieval.

The task construction follows a four‑stage expert‑driven workflow:

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

LLM‑based task drafting : an LLM generates a logically coherent, research‑ready draft from the core topic.

Multi‑expert review and revision : three experts audit design compliance, factual accuracy, logical coherence, and multimodal effectiveness.

Verification checklist creation : a checklist is compiled for systematic validation of report completeness.

TVIR‑Agent: Four‑Stage Hierarchical Multi‑Agent Pipeline

The TVIR‑Agent framework is a layered multi‑agent system built for text‑visual interleaved report generation. It consists of four core stages:

Planner : parses the user task, iteratively calls search and web‑scraping tools to retrieve relevant information, and synthesizes a structured outline. Each outline unit includes a chapter title and abstract, visual requirements, and research notes with citations, URLs, and key findings.

Visual Resource Instantiation : two specialized agents fulfill visual needs.

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

Chart Generator : retrieves data for distribution or relational content, verifies authenticity, generates Python plotting code, and executes it in a sandbox.

Writer : generates the report chapter by chapter. To maintain cross‑chapter coherence and reduce redundancy, it conditions generation on the current outline unit and a dynamically updated global context (titles, abstracts, sub‑structures of already generated chapters). It inserts research notes as evidence and calls search tools when evidence is insufficient. Visual descriptions determine insertion points, producing Markdown with interleaved text and images.

Polisher : at the document level, removes unreferenced citations, deduplicates URLs, renumbers references, updates in‑text citation markers, and reassigns sequential image IDs and labels throughout the report.

Dual‑Path Evaluation Framework

TVIR proposes a two‑component evaluation framework:

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

Visual Assessment (VA) : evaluates the relevance, correctness, and integration of visual elements.

Experimental Results

The research team evaluated nine deep‑research systems, including six commercial systems and three TVIR‑Agent variants.

Key Findings

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 performed best.

Variant Strengths :

TVIR‑Agent (GLM‑4.7) scored highest on TA, demonstrating strong textual ability.

TVIR‑Agent (Claude‑4.5‑Sonnet) dominated VA, showing superior visual support and cross‑modal alignment.

Citation Support Gap : TVIR‑Agent (GLM‑4.7) reached 68.64 on the Citation Support metric, 21.11 points higher than the best commercial system (Claude‑4.5‑Sonnet w/Search, 47.53).

Task Complexity Effects : As task difficulty increases, Instruction Alignment generally declines, while Analytical Depth & Breadth improve, indicating that complex tasks demand stronger multimodal coordination and long‑range reasoning.

Structural Error Analysis

Three error categories were examined:

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

Consistency errors : internal indexing failures such as missing or duplicate numbers.

Integrity errors : unavailable or incomplete referenced components (broken charts, missing captions).

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

Tool Usage Analysis

Different TVIR‑Agent variants displayed distinct tool‑usage patterns:

TVIR‑Agent (GLM‑4.7) used search and scraping tools most frequently, achieving the highest average effective citations (AEC) of 102.41. However, under a limited agent‑round budget, excessive retrieval reduced chart generation efficiency (average planned charts per task: 8.66, actually generated: 3.33, completion rate 38.45%).

TVIR‑Agent (Claude‑4.5‑Sonnet) adopted a more balanced strategy, maintaining a high AEC (86.14) while achieving the highest chart completion rate (94.61%).

These results indicate that a variant’s advantage depends not only on the underlying model but also on how it allocates tool usage between retrieval and chart generation.

Conclusion

TVIR introduces a unified benchmark and hierarchical agent framework for text‑visual interleaved report generation, a multimodal evaluation suite, and a comprehensive set of expert‑curated tasks. Extensive experiments show that TVIR‑Agent delivers strong overall performance, especially in evidence support and cross‑modal alignment, while revealing a persistent limitation of current deep‑research systems: they excel at textual synthesis but lag in visual integration. The authors hope TVIR will lay the groundwork for trustworthy multimodal deep‑research agents in the future.