YuhengSSS/VisionRL2-Gemma-4-12B

Paper: https://arxiv.org/abs/2609.19745

Vision-RL²: region-level policy optimization of the SD-RPN region proposal twig on a frozen Gemma-4-12B-it. This is the stage-2 (RL) checkpoint of the paper "Region-Level Policy Optimization for Fine-grained MLLM Perception".

The backbone weights are unchanged from google/gemma-4-12B-it; only the SD-RPN twig is trained.

Configuration

item	value
backbone	google/gemma-4-12B-it (frozen)
twig	K = 27, T = 3
stage	2 (region-level RL on top of the SD-RPN stage-1 twig)
placebo margin κ	1.0
weights	full model directory (backbone + twig), bfloat16 safetensors

How to use

These weights need the modeling code in YuHengsss/VisionRL2. They are not loadable for RoI inference through a plain AutoModel / AutoModelForCausalLM call: the RoI gating path (heatmap head, peak-relative gate, connected-component crop, sub-image splice) lives in the repository's model classes and evaluation harness.

git clone https://github.com/YuHengsss/VisionRL2
cd VisionRL2
hf download YuhengSSS/VisionRL2-Gemma-4-12B --local-dir output/rl/VisionRL2-Gemma-4-12B

# main table (generation + judge)
CHECKPOINT=output/rl/VisionRL2-Gemma-4-12B GPU_IDS=0,1 bash scripts/main_eval_gemma4.sh

# training-aligned protocol
CHECKPOINT=output/rl/VisionRL2-Gemma-4-12B bash scripts/aligned_eval_gemma4.sh

The checkpoint path is passed through the CHECKPOINT environment variable of scripts/main_eval*.sh and scripts/aligned_eval*.sh.

Links

• Code: https://github.com/YuHengsss/VisionRL2
• Data (RL pools + evidence maps): https://huggingface.co/datasets/YuhengSSS/VisionRL2-data
• Project page: https://yuhengsss.github.io/VisionRL2/

Citation

@article{shi2026visionrl2,
  title   = {Region-Level Policy Optimization for Fine-grained MLLM Perception},
  author  = {Shi, Yuheng and Pei, Xiaohuan and Dong, Minjing and Xu, Chang},
  journal = {arXiv preprint arXiv:2609.19745},
  year    = {2026}
}

@inproceedings{shi2026sdrpn,
  title     = {Catching the Details: Self-Distilled RoI Predictors for Fine-Grained MLLM Perception},
  author    = {Shi, Yuheng and Pei, Xiaohuan and Dong, Minjing and Xu, Chang},
  booktitle = {ICLR},
  year      = {2026}
}