JevNavAgent Pretrained Checkpoints (Qwen3-VL-2B-Instruct LoRA + Typed Decision Heads)

Official Stage-1 SFT and Stage-2 Trajectory-Level GRPO checkpoints for:

JevNavAgent: A Single-Forward Visual Navigation Agent with Trajectory-Level RL for Gigapixel Pathology
Yonghuang Wu, Zhenyang Liang, Pengfei Wang, Yize Ma, Guoqing Wu, and Jinhua Yu
School of Biomedical Engineering and Technological Innovation, Fudan University, Shanghai, China
Contact: yonghuangwu21@m.fudan.edu.cn

JevNavAgent Architecture and Two-Stage Training


Available Checkpoints

Each subdirectory contains lightweight PEFT LoRA weights (adapter/), the three factorized single-forward decision heads (heads.pt: evidence-sufficiency stopping, spatial-geometry candidate scoring, and gated contextual-bilinear answering), and jev_config.json (~73 MB per checkpoint).

Subdirectory Training Stage Primary Benchmark Performance
stage1_sft/ Stage-1 Counterfactual Trajectory Alignment SFT (formal_terra_v5) WSI-VQA 351: Step-$T$ 49.86% (Micro-Focal 49.31%, Clin. Aโ€“D 53.38%)
CAM16 Dev: 30.43% Loc Hit, 95.65% Spec, 71.74% Slide Acc
stage2_traj_grpo_vqa_kl/ Stage-2 Online Trajectory-Level GRPO (Joint VQA+KL) WSI-VQA 351: 51.57% Multi-Scale (50.43% Step-$T$), Micro-Focal 51.39%, Clin. Aโ€“D 54.73%, Net Zoom Gain +27 (42.74% -> 50.43%) at 5.63 s/episode
stage2_traj_grpo_cam16/ Stage-2 Online Trajectory-Level GRPO (CAMELYON16) CAM16 Official Test-129 (B=12): 22.45% Pos Loc Hit (91.7% stop conversion, 11/12 visited tumors), 93.75% Spec, 66.67% Strict Loc+Spec (67.44% Online), 69.77% Slide Acc

Zero-Shot Budget Scaling and Zoom Evidence Integration


Quick Usage

git clone https://github.com/yonghongwu/JevNavAgent.git
cd JevNavAgent && pip install -e .

# Download benchmark metadata & evaluation slides
jevnav-download --mode eval-slides --out ./data

# Evaluate Stage-2 VQA+KL checkpoint on Clean Visual WSI-VQA
jevnav-eval \
  --benchmark wsivqa_351 \
  --policy jevnav \
  --checkpoint renyimen/JevNavAgent:stage2_traj_grpo_vqa_kl \
  --backbone Qwen/Qwen3-VL-2B-Instruct \
  --data-dir ./data \
  --output results/jevnav_vqa.jsonl

# Evaluate Stage-2 CAM16 checkpoint on Verified CAMELYON16 Official Test-129
jevnav-eval \
  --benchmark cam16_test \
  --policy jevnav \
  --checkpoint renyimen/JevNavAgent:stage2_traj_grpo_cam16 \
  --backbone Qwen/Qwen3-VL-2B-Instruct \
  --data-dir ./data \
  --budget 12 \
  --output results/jevnav_cam16_test129.jsonl

Citation

@article{wu2026jevnavagent,
  title   = {JevNavAgent: A Single-Forward Visual Navigation Agent with Trajectory-Level RL for Gigapixel Pathology},
  author  = {Wu, Yonghuang and Liang, Zhenyang and Wang, Pengfei and Ma, Yize and Wu, Guoqing and Yu, Jinhua},
  journal = {arXiv preprint},
  year    = {2026},
  note    = {Checkpoints available at https://huggingface.co/renyimen/JevNavAgent}
}

License

Released under the Apache-2.0 License. Base backbone Qwen/Qwen3-VL-2B-Instruct remains subject to its original license.

Downloads last month
-
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for renyimen/JevNavAgent

Adapter
(131)
this model