UCAgents: Unidirectional Convergence for Visual Evidence Anchored Multi-Agent Medical Decision-Making

Qianhan Feng, Zhongzhen Huang, Yakun Zhu, Xiaofan Zhang, Qi DouarXiv 2025 · DOI · arXiv

Introduction

Vision-Language Models (VLMs) such as GPT-4 and LLaVA have demonstrated significant potential in medical diagnostics. However, these models often suffer from reasoning detachment, where they generate linguistically fluent explanations that lack grounding in the actual visual evidence of a medical image. While Multi-Agent Systems (MAS) simulating Multidisciplinary Teams (MDTs) have been proposed to mitigate this, they frequently fall victim to the information-noise paradox. As agents engage in open-ended debates, the accumulation of "textual noise" ($N_t$) can lead to rhetorical drift, where agents persuade one another through eloquent language rather than clinical evidence.

To address these challenges, we introduce UCAgents, a hierarchical framework that replaces circular debates with a structured, unidirectional flow designed to anchor reasoning in observable visual features.

Methodology: Unidirectional Convergence

UCAgents enforces a three-tier inquiry process that systematically reduces entropy and suppresses rhetorical noise. Unlike previous methods that allow for multiple rounds of position-changing debates, UCAgents limits interactions to targeted, evidence-based inquiries.

The UCAgents framework employs a three-tier hierarchical structure (Initial Diagnosis, Expert Review, and Critical Analysis) to ensure medical decisions are anchored in visual evidence.

The Three-Tier Architecture

  1. Tier 1: Initial Independent Diagnosis: Two agents perform independent analyses using temperature modulation ($ au=0.7$). This encourages the exploration of subtle visual variations without compromising the integrity of the medical image. Consensus leads to Tier 2, while disagreement escalates the case to Tier 3.
  2. Tier 2: Guidance Expert Review: A supervisor agent audits the Tier 1 consensus to identify "false consensus"—errors resulting from shared model biases. If a flaw is detected, the case is escalated.
  3. Tier 3: Critical Analysis & 1-Round Questioning: For complex cases, two "Critical Analysts" act as devil's advocates, tasked specifically with finding evidence against a proposed hypothesis. A Leader Agent then issues a single targeted inquiry to each analyst to resolve ambiguities before making a final arbitration.

Experimental Results

UCAgents was evaluated across four major medical VQA benchmarks: PathVQA, MIMIC-CXR-VQA, VQA-RAD, and SLAKE-VQA. The framework achieved state-of-the-art accuracy, notably reaching 71.3% on PathVQA, a 6.0% improvement over previous benchmarks.

Visual Evidence and Efficiency

A key metric for UCAgents is its ability to resolve reasoning detachment. The system successfully recalled 79.2% of critical visual evidence, a significant improvement over the 35.8% achieved by the leading multi-agent competitor, MDAgents.

Comparison of visual evidence coverage (recall) between single-agent, UCAgents, and previous multi-agent (MDAgents) methods, highlighting the reduction in missing evidence.

Furthermore, UCAgents demonstrates remarkable computational efficiency. By eliminating redundant discussion rounds, it reduced token usage by 87.7% compared to MDAgents. On open-source models like Qwen2.5VL-7B, the system consumed only 14.8% of the tokens required by traditional debate-based methods, making it highly suitable for real-world clinical deployment.

Conclusion

UCAgents effectively solves the information-noise paradox in medical AI by prioritizing adversarial auditing over open debate. By forcing agents to anchor their arguments in visual features and limiting the rounds of interaction, the framework provides more stable, accurate, and cost-effective diagnostic support. Future work will focus on incorporating "epistemic humility," allowing the system to explicitly identify cases where visual evidence is insufficient for a definitive conclusion.

Summary updated

BibTeX
@misc{feng2025ucagents,
  author = {Qianhan Feng and Zhongzhen Huang and Yakun Zhu and Xiaofan Zhang and Qi Dou},
  title = {{UCAgents: Unidirectional Convergence for Visual Evidence Anchored Multi-Agent Medical Decision-Making}},
  year = {2025},
  eprint = {2512.02485},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2512.02485},
}
Figure