Chi Phan*, Tianyi Zhang*, Yufeng Wu, Qiaochu Xue, Jiajie Zhang, Linghan Cai, Zeyu Liu, Sudong Wang, Yueming Jin, Dan Hu
Overview of the proposed cross-scale benchmark construction and model optimization framework. (A) Expert-verified diagnostic paths link clinically relevant 10×, 40×, and 200× ROIs from the same WSI, providing scale-specific captions and cross-scale evidence anchors. (B) From these paths, we construct cross-scale semantic reasoning and visual grounding tasks, with adversarial text-only screening to reduce language shortcuts and structure-controlled distractor sampling to reduce superficial visual shortcuts. (C) PathScale-R1 is optimized by difficulty- driven reasoning distillation followed by reinforcement learning with accuracy, format, and scale-aware reasoning structure rewards.
Dataset statistics and benchmark performance. (A) PathScale-VQA component statistics and organ distribution. (B) Task-wise performance of representative VLMs across single-scale and proposed cross-scale VQA benchmark.
Training Dynamics:
PathScale-R1/
├── preprocess/
│ ├── generate_vqa_data/ # Cross-scale VQA generation and split creation
│ └── prompts/ # Prompt templates and task constraints
├── script/
│ ├── preprocess/ # Preprocessing entrypoints
│ ├── train/ # SFT and GRPO training scripts
│ └── postprocess/ # Checkpoint export / merge scripts
├── LLaMA-Factory/ # SFT framework
├── verl/ # RL framework
└── README.md
git clone [repo url placeholder]
cd PathScale-R1
cp script/.env.example script/.env
# Edit script/.env with your local paths and API keysConfigure script/.env:
# Data processing
DATA_DIR=/path/to/triplet_raw_data
ROOT=/path/to/PathScale-R1
PROCESSED_DIR=/path/to/processed_data
# Training
ACTOR_MODEL_DIR=/path/to/base_pathor1_model
LOG_DIR=/path/to/logs
WANDB_DIR=/path/to/wandb_logs
RESULTS_DIR=/path/to/results_dir
# Exporting checkpoints
RAW_CKPT_DIR=/path/to/raw_ckpt_dir
OUR_MODEL_DIR=/path/to/our_final_model_dir
# Secrets
GEMINI_API_KEY=
OPENAI_API_KEY=
DASHSCOPE_API_KEY=
HF_TOKEN=This repo currently uses two training stacks:
verlfor GRPO-based RL trainingLLaMA-Factoryfor SFT
Suggested setup:
conda create -n verl python=3.10 -y
conda activate verl
pip install -e verl/
conda create -n sft python=3.10 -y
conda activate sft
pip install -e LLaMA-Factory/CUDA / PyTorch version requirements should be finalized against the exact training environment used for release.
More details on code implementation are coming soon!
This work was supported by the Ministry of Education, Singapore, under the Tier 1 grant (24-1250-P0001) and Tier 2 grant (T2EP20224-0028), and by PuzzleLogic Pte Ltd, Singapore.
We gratefully acknowledge the open-source projects that made the development of PathScale-R1 possible:
- verl, for the reinforcement learning training framework.
- LLaMA-Factory, for the unified fine-tuning pipelines.
- vLLM, for efficient large language model inference and serving.
We also acknowledge the following open-source models used for comparison in our experiments: Qwen2.5-VL-7B, Qwen3-VL-8B, InternVL3.5-8B, MiMo-VL-7B, LLaVA-Med-7B, HuatuoGPT-Vision-7B, MedVLThinker-7B, QoQ-Med-VL-7B, Lingshu-7B, HealthGPT-8B, OctoMed-7B, Quilt-LLaVA, CLOVER, and Patho-R1.
We sincerely thank the developers and contributors of these projects for their excellent work and for making their code and models publicly available to the research community.
If you find our work helpful, please consider citing our paper and the frameworks we build upon:
@article{phan2026pathscale,
title={PathScale-R1: Cross-scale Reasonin
5291
g for Pathological Image Analysis},
author={Phan, Chi and Zhang, Tianyi and Wu, Yufeng and Xue, Qiaochu and Zhang, Jiajie and Cai, Linghan and Liu, Zeyu and Wang, Sudong and Jin, Yueming and Hu, Dan},
journal={arXiv preprint arXiv:2607.23794},
year={2026}
}