MedVerse: Efficient and Reliable Medical Reasoning via DAG-Structured Parallel Execution

Bridging the gap between medical reasoning quality and inference efficiency through DAG-structured parallel execution.

🔥 News

• [04/15/2026] The code of MedVerse was released!
• [02/10/2026] MedVerse paper released on arXiv.

📖 Overview

Medical reasoning is inherently multi-faceted: answering a clinical question requires simultaneously reasoning over differential diagnoses, laboratory findings, drug interactions, and treatment guidelines. Standard autoregressive LLMs collapse these parallel cognitive tasks into a single sequential chain-of-thought — forcing steps that are logically independent to wait on one another.

MedVerse reformulates medical inference as a parallelizable directed acyclic graph (DAG) grounded in Petri net theory. The framework has three components:

Component	What it does
MedVerse Curator	Automated pipeline that synthesizes knowledge-grounded medical reasoning and converts it into Petri net DAG structures for training
Topology-Aware Attention	Training-time attention mechanism with adaptive position indices that enables the model to reason across parallel branches while maintaining logical coherence
MedVerse Inference Engine	Customized SGLang-based server that executes the DAG at inference time — forking independent steps into concurrent GPU requests and joining their outputs

Together, these yield a model that matches specialized medical models while improving general-purpose LLMs by up to 8.9%, and an inference engine that reduces latency by 1.3× and increases generation throughput by 1.7× through parallel decoding.

🤖 Key Features

• DAG-structured parallel inference: MedVerse models emit a <Plan> block encoding explicit step dependencies as a Petri net. The inference engine parses this into a DAG, identifies independent reasoning paths, and dispatches them as concurrent GPU requests — no client changes required.
• Topology-aware attention with adaptive position indices: The fine-tuned model learns to produce coherent parallel reasoning branches through a topology-aware attention mask applied during training. Adaptive position indices ensure each branch maintains positional context relative to the shared prefix, not to each other.
• Knowledge-grounded medical reasoning (MedVerse Curator): Training data is synthesized by the MedVerse Curator: a pipeline that decomposes medical questions into multi-step reasoning graphs, grounded in clinical knowledge sources, then converts them into the Petri net format used at inference time.
• Radix-cache prefix sharing: All parallel child requests in Phase II share the same Phase I KV-cache prefix via SGLang's radix attention tree, making Phase II prefill cost near-zero regardless of the number of parallel branches.

---

🚀 Getting Started

Installation

git clone https://github.com/aiming-lab/MedVerse.git

Training (medverse):

cd MedVerse/train
conda create -n medverse python=3.10 -y
conda activate medverse
pip install -r requirements.txt

Inference Engine (medverse-engine):

cd MedVerse/MedVerse-Engine
conda create -n medverse-engine python=3.11 -y
conda activate medverse-engine
bash install.sh

---

🏋️ Training

MedVerse is fine-tuned from Qwen2.5-7B-Instruct and LLaMA-3.1-8B-Instruct using topology-aware attention on the MedVerse14k dataset — 13,904 medical questions annotated with knowledge-grounded DAG reasoning paths generated by the MedVerse Curator. The training dataset is available on 🤗 HuggingFace.

Data Preparation

The scripts below automatically download MedVerse14k from HuggingFace and convert it into the training format for each model. Run the one matching your target model:

cd data

# Qwen2.5 — converts to ChatML format, saves to data/datasets/MedVerse14k
python preparation/prepare_train.py

# LLaMA-3 — converts to LLaMA chat format, saves to data/datasets/MedVerse14k-LLaMA
python preparation/prepare_train_llama.py

cd ..

Or generate the dataset from scratch using the MedVerse Curator pipeline — see data/README.md for the full data generation guide including input format and step-by-step instructions.

Train MedVerse (Topology-Aware Attention SFT)

bash train/scripts/launch_train_qwen.sh   # Qwen2.5-7B-Instruct
bash train/scripts/launch_train_llama.sh  # LLaMA-3.1-8B-Instruct

See train/README.md for configuration details.

---

⚡ Inference Engine

Start the Server

python -m sglang.srt.entrypoints.medverse_server \
    --model-path /path/to/MedVerse-Qwen2.5-7B \
    --tp-size 1 \
    --port 30000 \
    --trust-remote-code \
    --mem-fraction-static 0.85

Wait for Server is ready in the logs.

Try an Example

cd MedVerse/MedVerse-Engine/example

python example.py \
    --server_url http://localhost:30000 \
    --prompts_dir ./prompt

See MedVerse-Engine/README.md for full installation and configuration details.

---

📚 Citation

If you find our work helpful, please consider citing:

@article{chen2026medverse,
    title         = {MedVerse: Efficient and Reliable Medical Reasoning via DAG-Structured Parallel Execution},
    author        = {Jianwen Chen and Xinyu Yang and Peng Xia and Arian Azarang and Yueh Z Lee and Gang Li and Hongtu Zhu and Yun Li and Beidi Chen and Huaxiu Yao},
    journal       = {arXiv preprint arXiv:2602.07529},
    year          = {2026},
    url           = {https://arxiv.org/abs/2602.07529},
}

---

🙏 Acknowledgement

We would like to express our gratitude to the open-source community and the following projects for making this work possible: SGLang, Multiverse Engine, Qwen, MedReason, etc.