YuhengSSS/SDRPN-Qwen3.5-9B

Papers: https://arxiv.org/abs/2609.19745 (Vision-RL²) · https://arxiv.org/abs/2509.16944 (SD-RPN)

SD-RPN stage-1 checkpoint: a self-distilled RoI predictor twig (K = 21, T = 3) trained on a frozen Qwen/Qwen3.5-9B. This is the initialisation of the Vision-RL² RL run YuhengSSS/VisionRL2-Qwen3.5-9B.

The backbone weights are unchanged from the base model; only the three attached twig blocks are trained, from self-distilled attention pseudo-labels (no human RoI annotation).

Configuration

item	value
backbone	`Qwen/Qwen3.5-9B` (frozen)
twig	K = 21, T = 3
stage	1 (SD-RPN self-distilled pseudo-label training)
weights	full model directory (backbone + twig), bfloat16 safetensors
next stage	region-level RL → `YuhengSSS/VisionRL2-Qwen3.5-9B`

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/SDRPN-Qwen3.5-9B --local-dir output/sdrpn/qwen3_5-9b-sdrpn-K21T3

# stage 2: region-level RL starting from this checkpoint
PHASE_A_CKPT=output/sdrpn/qwen3_5-9b-sdrpn-K21T3 DATASET_ROOT=datasets \
  bash scripts/train_rl_qwen3_5_9b.sh

# evaluate the stage-1 checkpoint itself (training-aligned protocol)
MODEL=qwen3_5 CHECKPOINT=output/sdrpn/qwen3_5-9b-sdrpn-K21T3 CAP=576 bash scripts/aligned_eval.sh

Expected numbers (sanity check after download)

The Vision-RL² README publishes the per-benchmark SD-RPN (stage 1) row for Qwen3.5-4B and Gemma-4-12B-it; for the 9B backbone the paper reports the stage-2 model only, so use the project-page summary figure as the reference point: the Qwen3.5-9B Vision-RL² main-table average is 80.1, and this stage-1 checkpoint is the initialisation that RL run starts from. A cheap structural check after download: config.json carries twig_K: 21 / twig_T: 3, and the model.safetensors total is within a few hundred MB of the base Qwen/Qwen3.5-9B (~20 GB) because only the twig differs.

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}
}