dmis-lab / InternVL3-8B-Instruct-MRPO

InternVL3-8B-Instruct-MRPO

MRPO is a novel reinforcement learning framework that improves medical multimodal reasoning by directly addressing failures in the reasoning process. It reshapes GRPO-style advantages using both answer-level and step-wise process rewards, assigning exponentially larger penalties to earlier invalid steps when the final answer is incorrect, thereby correcting early-stage failures before they cascade while preserving successful trajectories. By redistributing the learning signal according to where reasoning first fails, MRPO induces transferable reasoning that improves both reasoning quality and final answer accuracy across diverse medical VQA benchmarks.

Code: github

Project Page: page

Paper: Breaking Failure Cascades: Step-Aware Reinforcement Learning for Medical Multimodal Reasoning

Quick Start

import torch
import torchvision.transforms as T
from PIL import Image
from torchvision.transforms.functional import InterpolationMode
from transformers import AutoModel, AutoTokenizer

# Load the model (MRPO InternVL3 checkpoint; or a local trained checkpoint path)
model_path = "dmis-lab/InternVL3-8B-Instruct-MRPO"
model = AutoModel.from_pretrained(
    model_path,
    torch_dtype=torch.bfloat16,
    device_map="auto",
    trust_remote_code=True,
).eval()
tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)


# InternVL3 image preprocessing: dynamic 448x448 tiling (up to 12 tiles + thumbnail)
def load_image(image_path, input_size=448, max_num=12):
    image = Image.open(image_path).convert("RGB")
    w, h = image.size
    ratios = sorted(
        {(i, j) for n in range(1, max_num + 1) for i in range(1, n + 1) for j in range(1, n + 1) if i * j <= max_num},
        key=lambda r: r[0] * r[1],
    )
    best, best_diff = (1, 1), float("inf")
    for r in ratios:
        diff = abs(w / h - r[0] / r[1])
        if diff < best_diff or (diff == best_diff and w * h > 0.5 * input_size * input_size * r[0] * r[1]):
            best, best_diff = r, diff
    tw, th = input_size * best[0], input_size * best[1]
    resized = image.resize((tw, th))
    tiles = [
        resized.crop((x * input_size, y * input_size, (x + 1) * input_size, (y + 1) * input_size))
        for y in range(best[1]) for x in range(best[0])
    ]
    if len(tiles) > 1:
        tiles.append(image.resize((input_size, input_size)))
    transform = T.Compose([
        T.Resize((input_size, input_size), interpolation=InterpolationMode.BICUBIC),
        T.ToTensor(),
        T.Normalize(mean=(0.485, 0.456, 0.406), std=(0.229, 0.224, 0.225)),
    ])
    return torch.stack([transform(t) for t in tiles])


# Example usage (InternVL3 default system prompt is applied inside model.chat)
image_path = "path/to/medical/image.jpg"
question = "What can you see in this medical image?"

question_text = (
    f"<image>\n{question} Think step-by-step and enclose your reasoning in "
    "<think>...</think> tags. Then provide your answer in <answer>...</answer> tags."
)
pixel_values = load_image(image_path).to(torch.bfloat16).to(model.device)

# Inference (greedy decoding, matching inference.py)
output_text = model.chat(tokenizer, pixel_values, question_text, dict(max_new_tokens=512, do_sample=False))
print(output_text)

Citation

@misc{jung2026breakingfailurecascadesstepaware,
      title={Breaking Failure Cascades: Step-Aware Reinforcement Learning for Medical Multimodal Reasoning}, 
      author={Junha Jung and Minbyul Jeong and Suhyeon Lim and Sungwook Jung and Jaehoon Yun and Taeyun Roh and Mujeen Sung and Jaewoo Kang},
      year={2026},
      eprint={2606.31825},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2606.31825}, 
}

License

This model is released under the Apache 2.0 license.