Technical Summary: Ryze – Evidence-Enriched Data Synthesis from Biomedical Papers

Problem Statement

General-purpose Vision-Language Models (VLMs) currently lack the reliability required for biomedical research workflows. Scientific reasoning in this domain depends heavily on interpreting evidence distributed across multiple modalities within a paper: figures, charts, tables, captions, and the surrounding explanatory prose. Existing post-training pipelines face two primary bottlenecks:

1. Data Scarcity and Cost: Acquiring high-quality, domain-specific QA pairs typically requires expensive expert annotation.
2. Evidence Degradation: Existing synthetic data methods often fail to preserve the structural context (e.g., the link between a chart axis and its caption, or a table row and its header) necessary for valid scientific answers. When training examples drop these components, models tend to learn shallow cues or memorized patterns rather than evidence-grounded reasoning. Furthermore, general-purpose OCR and document conversion tools often introduce hallucinations or layout errors when processing complex scientific PDFs, which propagate into training data and degrade model performance.

Methodology

Ryze is a fully automated system designed to convert raw biomedical papers into an evidence-enriched training set and a specialized VLM (BioVLM-8B) without human annotation. The system operates through four interconnected modules in a closed loop:

1. Chart-Aware Extraction and Cleansing

To mitigate OCR hallucinations and layout errors, Ryze employs a five-stage pipeline:

• Layout Detection: Uses Surya to segment pages into text blocks, figures, tables, and captions, preserving spatial relationships often lost in single-pass OCR.
• Structured Conversion: Converts text regions to Markdown while maintaining section hierarchy.
• Cross-Reference Repair: Recovers inline references (e.g., "Figure 3") and binds visual elements with their captions and interpreting prose, ensuring the full evidential context is available.
• Specialized Extraction: Applies chart/table-aware extraction (using GLM-OCR) to parse charts into structured numerical descriptions and tables into HTML with correct handling of merged cells and multi-row headers.
• LLM-Based Cleansing: A final pass (using Qwen3) performs hallucination detection, domain-terminology repair (e.g., correcting chemical formulae), and cross-element consistency verification before data enters the synthesis pipeline.

2. Evidence-Enriched QA Synthesis

The synthesis pipeline transforms structured documents into high-quality QA pairs:

• Task-Aware Question Generation: Instead of copying benchmark questions, the system extracts coarse-grained reasoning skill categories (e.g., "chart interpretation") from evaluation benchmarks. These categories are diversified by a large VLM (Qwen3-VL-235B) to generate new questions. Crucially, no specific benchmark questions or answers are used in training; all answers are grounded exclusively in the source PDF corpus.
• Evidence-Retrieval: Each generated question is paired with a comprehensive evidence set retrieved from the structured document store, including captions, OCR annotations, and related paragraphs.
• Active Data Augmentation: An evaluation feedback loop identifies underperforming domain categories, triggering the automatic search and processing of new open-access papers to synthesize targeted QA pairs.

3. Progress-Gated Post-Training

Ryze adopts a two-phase training strategy to optimize data efficiency and reasoning capabilities:

• Calibration Phase (SFT): The base model (Qwen3-VL-8B) undergoes Supervised Fine-Tuning (SFT) with LoRA. Training alternates between text-only and vision QA batches to inject domain terminology and visual-textual alignment.
• Progress Gate: The system monitors benchmark performance after every ~1M-token increment. When accuracy improvements stagnate, SFT is halted.
• Refinement Phase (RL): The accumulated data is converted into a Reinforcement Learning (RL) format using Group Relative Policy Optimization (GRPO). This phase strengthens evidence-grounded reasoning without requiring a separate reward model, focusing on generating coherent reasoning chains for complex biomedical questions.

4. Evaluation and Feedback

The system evaluates performance across benchmark categories (e.g., LitQA2, FigQA, TableQA). Weaknesses detected below a configurable threshold update the QA template library, triggering a new round of data synthesis to iteratively refine the model.

Key Results

Evaluated on LAB-Bench (1,967 samples across eight biology categories), the resulting BioVLM-8B model demonstrates significant improvements:

• Performance: BioVLM-8B achieves a 48.0% weighted accuracy, outperforming the base Qwen3-VL-8B by +12.6 percentage points (pp) and surpassing the commercial GPT-5.2 by +3.8 pp.
• Training Efficiency: The SFT-only version (43.7%) already matches GPT-5.2 (44.2%). The subsequent GRPO stage provides the critical +4.3 pp margin for overall superiority.
• Data Quality: Under an equal token budget, Ryze's synthesized data outperforms human-curated datasets PubMedQA (+17.1 pp) and MedQA (+14.7 pp).
• Cost Efficiency: The entire end-to-end pipeline (including OCR, synthesis, and training) costs under $200 on cloud GPUs, making it accessible to small labs and reproducible without expert annotation.
• Generalization: The evidence-enriched dataset improves performance across different base architectures (Qwen2.5, LLaMA-3.2, Gemma-2), with the largest gains observed on models with both text and vision components.

Significance and Claims

The paper claims that Ryze addresses a critical systems problem: how to transform raw scientific literature into specialized VLMs at low cost while preserving the "evidence-grounded" nature of scientific reasoning. The core insight is that preserving the full contextual structure of scientific documents (visual elements, captions, and referring text) produces training data fundamentally superior to both human-curated alternatives and context-agnostic automated approaches.

By shifting domain adaptation from an "expert-annotation bottleneck" to a reproducible, automated pipeline, Ryze enables small research teams to produce models that outperform large, expensive commercial models on specific scientific benchmarks. The authors emphasize that while the current focus is biomedicine, the system is designed to generalize to other scientific fields, with early experiments showing promise in climate change, geoscience, and civil engineering.

Limitations

The authors note that the evaluation is currently limited to the biology domain, and generalization to other fields requires further validation. Additionally, GPT-5.2 retains advantages in categories involving heavy visual understanding (FigQA) and sequence analysis (SeqQA, SuppQA), indicating room for improvement in multimodal reasoning. The scaling behavior of the progress-gated paradigm with larger models also warrants further investigation.