Title: Ryze: Evidence-Enriched Data Synthesis from Biomedical Papers

Abstract:

General-purpose Vision-Language Models (VLMs) often prove unreliable in biomedical research contexts, as accurate responses in scientific literature frequently require synthesizing evidence distributed across various components such as figures, tables, charts, captions, and associated text. Current post-training methodologies are hindered by two primary issues: the high expense of expert annotation and the loss of critical evidence structure inherent in synthetic datasets. To address these challenges, we introduce Ryze, an automated framework designed to transform raw biomedical papers into both a domain-specialized VLM and a training dataset enriched with evidence.

Ryze generates question-and-answer pairs that include comprehensive supporting materials, such as visual elements, captions, extracted structures, and relevant paragraphs. The system mitigates layout and Optical Character Recognition (OCR) errors through chart- and table-aware extraction techniques, alongside LLM-based data cleansing. Furthermore, it employs a progress-gated post-training approach that integrates supervised fine-tuning with reinforcement learning.

Initiating with the Qwen3-VL-8B foundation, Ryze creates BioVLM-8B at a cost of less than USD 200. This model achieves a weighted accuracy of 48.0% on the LAB-Bench benchmark, marking a 12.6 percentage point improvement over the base model and outperforming GPT-5.2 by 3.8 percentage points. We are making both the Ryze system and the trained BioVLM-8B model available as open-source resources.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC