E-MRL: Cross-view Aligned Evidence-driven Multimodal Reinforcement Learning for Reliable 3D Tumor Analysis

Researchers have proposed a new reinforcement learning framework, E-MRL, designed to improve the reliability of 3D tumor analysis by reducing visual hallucinations in Vision-Language Models, according to a paper posted on arXiv [1]. The framework, formally named cross-view aligned Evidence-driven Multimodal Reinforcement Learning, addresses a critical weakness in current Vision-Language Models (VLMs) used for medical imaging. While VLMs can generate volumetric medical reports, they often produce visual hallucinations and lack grounding in the actual 3D CT data [1]. Standard training methods, including Supervised Fine-Tuning and Reinforcement Learning, typically reward text fidelity, which can lead to correct diagnoses based on language patterns rather than genuine visual perception [1]. To counter this, the authors structure the report generation as a Markov Decision Process with three stages: diagnosis, localization, and verification [1]. The model is explicitly trained to identify a "key evidence slice" alongside the global diagnostic report, grounding its findings in verifiable visual evidence [1]. A novel cross-view consistency reward validates the semantic alignment between the standard report and a local visual re-query of the selected key slice, providing additional rewards for correctly-localized reasoning [1]. The paper, submitted to arXiv's Electrical Engineering and Systems Science section on June 22, 2026, reports that experiments on large-scale 3D CT tumor datasets show E-MRL significantly reduces hallucinations and improves diagnostic accuracy compared to existing baselines [1]. The authors present it as a clinically interpretable solution for visually-grounded tumor analysis [1]. arXiv, which began on August 14, 1991, is an open-access repository for electronic preprints that are moderated but not peer-reviewed [6]. As of November 2024, the submission rate was about 24,000 articles per month, and the repository surpassed two million articles by the end of 2021 [6]. The platform also hosts arXivLabs, a framework enabling community collaborators to develop and share experimental tools directly on the website, such as the Bibliographic Explorer and CORE Recommender [4][5].