ReGround Qwen2.5-VL-7B

This is the final model checkpoint for ReGround (Project Page, GitHub), a visual reasoning method that teaches a vision-language model to request fresh visual evidence through a structured <reground> self-diagnosis.

When <reground> is emitted, the inference controller should re-inject the original image and ask the model to complete its reasoning and final answer. The public repository provides the two-round inference adapter, data construction pipeline, SFT recipe, and Stage-2 GRPO implementation.

The released weights were trained in two stages: full-parameter supervised fine-tuning on structured ReGround trajectories, followed by GRPO policy optimization of the re-examination behavior.

Usage

from transformers import AutoProcessor, Qwen2_5_VLForConditionalGeneration

model_id = "SESPOIR/ReGround-Qwen2.5-VL-7B"
model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
    model_id,
    dtype="auto",
    device_map="auto",
)
processor = AutoProcessor.from_pretrained(model_id)

For complete two-round visual inference and evaluation, follow the ReGround repository.

Training

Stage 1 applies full-parameter supervised fine-tuning to Qwen2.5-VL-7B-Instruct using structured visual reasoning trajectories. Stage 2 uses GRPO to optimize self-diagnosis, answer accuracy, and output format. The training format uses <think>, <reground>, and <answer> blocks; portable configurations for both stages are available in the project repository.

Limitations

The model can produce incorrect reasoning or answers and may trigger visual re-examination unnecessarily. It is intended for research use and should not be relied on for high-stakes decisions without independent verification.