MedViz: An Agent-based, Visual-guided Research Assistant for Navigating Biomedical Literature

Huan He, Xueqing Peng, Yutong Xie, Qijia Liu, Chia-Hsuan Chang, Lingfei Qian, Brian Ondov, Qiaozhu Mei, Hua Xu

Abstract

Biomedical researchers face increasing challenges in navigating millions of publications in diverse domains. Traditional search engines typically return articles as ranked text lists, offering little support for global exploration or in-depth analysis. Although recent advances in generative AI and large language models have shown promise in tasks such as summarization, extraction, and question answering, their dialog-based implementations are poorly integrated with literature search workflows. To address this gap, we introduce MedViz, a visual analytics system that integrates multiple AI agents with interactive visualization to support the exploration of the large-scale biomedical literature. MedViz combines a semantic map of millions of articles with agent-driven functions for querying, summarizing, and hypothesis generation, allowing researchers to iteratively refine questions, identify trends, and uncover hidden connections. By bridging intelligent agents with interactive visualization, MedViz transforms biomedical literature search into a dynamic, exploratory process that accelerates knowledge discovery.

Introduction

The biomedical field is experiencing an information explosion, with millions of publications spanning increasingly specialized domains. Traditional search tools like PubMed are effective for targeted retrieval but often fail to provide a holistic view of a research landscape. Researchers struggle with "sensemaking"—the process of understanding the global structure of a field, tracking the evolution of topics, and identifying underexplored "white spaces." While Large Language Models (LLMs) offer powerful summarization capabilities, they often function as "black boxes," providing answers without clear context or visibility into the underlying literature. MedViz addresses these challenges by integrating a large-scale semantic map with a multi-agent AI framework, enabling grounded, visual-guided exploration.

System Architecture

The MedViz architecture is built upon three primary layers designed to handle massive datasets and complex reasoning tasks.

Data Processing Pipeline

The system begins by ingesting bibliographic metadata and abstracts from sources like PubMed and iCite. The pipeline transforms raw data into an interactive visualization through several stages:

  • Text Embedding: A pretrained LLM (Gemma-300m) generates high-dimensional semantic embeddings for each paper.
  • Dimension Reduction: The LargeVis algorithm projects these embeddings into 2D coordinates while preserving local semantic neighborhoods.
  • Clustering and Labeling: HDBSCAN identifies hierarchical clusters, which are then labeled by LLMs to provide human-readable topic descriptions.
The MedViz data processing pipeline, illustrating the transition from raw metadata to high-dimensional embeddings and finally to a 2D interactive visualization.

Visual Analytics Interface

The front-end, built with WebGL and Three.js, renders up to 1.2 million papers as a point cloud. To manage visual complexity, the system employs the Hammer Bundle algorithm for "edge bundling," allowing citation networks to be visualized without overwhelming the user. The interface provides comprehensive tools for search, selection, and temporal analysis.

The MedViz user interface, showcasing the semantic map, selection tools, and integrated analysis panels for large-scale literature exploration.

Multi-agent Reasoning Framework

The "brain" of MedViz is a multi-agent system orchestrated by a Scholar Agent. This agent interprets user queries and delegates specialized tasks:

  • Evidence Agent: Extracts structured data and summarizes outcomes from specific paper selections.
  • Analytical Agent: Identifies statistical trends and performs comparative analyses.
  • Discovery Agent: Assists in hypothesis generation and identifying research gaps.

Key Insights and Features

MedViz introduces several innovations in literature navigation:

  • Space-based Sensemaking: By representing similarity through spatial proximity, researchers can visually grasp the boundaries and density of different research areas.
  • Visual Grounding: Unlike standard RAG systems, MedViz agents are "grounded" in the user's visual selection. The AI only analyzes papers within a user-defined lasso or cluster, ensuring transparency and control.
  • Temporal Stability: The system allows users to filter by publication year without altering the layout, enabling stable observation of how topics emerge and migrate over decades.
  • Closed-loop Interaction: Agents can interact directly with the UI, such as highlighting specific papers on the map in response to a query.

Results and Discussion

MedViz demonstrates high scalability, maintaining responsive performance for up to 1.2 million points in a standard web browser. This is achieved through spatial indexing (Quadtrees) and GPU-accelerated rendering. The system's versatility has been validated through workflows ranging from rapid field familiarization to conducting "mini systematic reviews."

Despite its strengths, the system faces limitations. Browser memory constraints currently cap the point count at 1.2 million, necessitating future "level-of-detail" strategies similar to geographic mapping tools. Additionally, while the multi-agent framework is robust, future work will focus on adding verification loops to further mitigate LLM hallucinations and expanding data "on-ramps" for third-party reference managers.

Conclusion

MedViz represents a significant step toward bridging the gap between intelligent agents and interactive visualization. By transforming the literature search from a list-based retrieval task into a dynamic, exploratory process, MedViz empowers researchers to navigate the vast biomedical landscape with greater clarity and insight.

Summary updated

Abstract

Biomedical researchers face increasing challenges in navigating millions of publications in diverse domains. Traditional search engines typically return articles as ranked text lists, offering little support for global exploration or in-depth analysis. Although recent advances in generative AI and large language models have shown promise in tasks such as summarization, extraction, and question answering, their dialog-based implementations are poorly integrated with literature search workflows. To address this gap, we introduce MedViz, a visual analytics system that integrates multiple AI agents with interactive visualization to support the exploration of the large-scale biomedical literature. MedViz combines a semantic map of millions of articles with agent-driven functions for querying, summarizing, and hypothesis generation, allowing researchers to iteratively refine questions, identify trends, and uncover hidden connections. By bridging intelligent agents with interactive visualization, MedViz transforms biomedical literature search into a dynamic, exploratory process that accelerates knowledge discovery.
BibTeX
@misc{he2026medviz,
  author = {Huan He and Xueqing Peng and Yutong Xie and Qijia Liu and Chia-Hsuan Chang and Lingfei Qian and Brian Ondov and Qiaozhu Mei and Hua Xu},
  title = {{MedViz: An Agent-based, Visual-guided Research Assistant for Navigating Biomedical Literature}},
  year = {2026},
  eprint = {2601.20709},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2601.20709},
}
Figure