๐ Paper | ๐ Project | ๐ค ZwZ2-RL-VQA | ๐ค ZwZ2-4B | ๐ค ZwZ2-9B | ๐ค ZwZ2-27B
Zooming without Zooming v2 connects fine-grained visual evidence across regions. We train ZwZ2-4B/9B/27B with reinforcement learning on 20.4K verified, multi-region VQA examples. Training uses full images; inference needs no zooming tools. The models transfer across perception, spatial reasoning, long-context understanding and multimodal agents.
EvidenceConnect turns evidence from 3โ7 image regions into verifiable supervision:
- Discover grounded regions and synthesize questions that connect their evidence.
- Verify answers through cross-model agreement and adaptive region-ablation tests.
- Train on the full-image view with GRPO, without supervised fine-tuning.
Results on 25 benchmarks across six domains (accuracy, %).
| Model | Base model |
|---|---|
| ZwZ2-4B | Qwen3.5-4B |
| ZwZ2-9B | Qwen3.5-9B |
| ZwZ2-27B | Qwen3.6-27B |
All three scales use the same 20.4K training examples, synthesized from SA-1B images with cross-model verification. Privileged crops are used during data construction only; the policy learns from the full-image view.
| Component | Directory |
|---|---|
| Data synthesis | data_synthesis/ |
| Training | verl/ |
| Evaluation | eval/ |
cd data_synthesis
cp configs/local_env.sh.example configs/local_env.sh # then edit keys/model names
pip install openai httpx pillow pyarrow pytestSet your API URL, keys and model names in configs/local_env.sh (gitignored).
The only input is a folder of images:
# 1) Stage 01 โ scan the image folder, build a manifest, and drop
# low-content images (resolution / entropy / edge density / std-dev):
python scripts/01_filter_manifest.py \
--image_folder /path/to/images \
--output_jsonl data/filtered_manifest.jsonl
# 2) Continuous synthesis over the filtered manifest:
bash run_pipeline_continuous.sh [TARGET_VALID] [BATCH_SIZE] [MAX_BATCHES] [MANIFEST]
# MANIFEST defaults to data/filtered_manifest.jsonl (the stage-01 output)The synthesized records are exported to a verl-ready parquet for RL training:
python scripts/05_export_dataset.py \
--validated_jsonl data/all_valid.jsonl \
--output_parquet mcq_train.parquet \
--format mcq--format mcq is the multiple-choice format consumed by the training
launchers below. Any verl-format validation parquet can be used as the
evaluation set.
pytest -q tests/Train with GRPO using verl. Choose the launcher for your base model:
| Model | Script |
|---|---|
| Qwen3.5-4B | verl/train_zwz2_qwen3_5_4b.sh |
| Qwen3.5-9B | verl/train_zwz2_qwen3_5_9b.sh |
| Qwen3.6-27B | verl/train_zwz2_qwen3_6_27b.sh |
pip install -e ./verl # install the vendored verl frameworkSet the model and dataset paths:
cd verl
MODEL_PATH=/path/to/Qwen3.5-9B \
TRAIN_FILE=/path/to/mcq_train.parquet \
TEST_FILE=/path/to/val.parquet \
bash train_zwz2_qwen3_5_9b.shRun the same way for the other sizes; checkpoints, logs and rollout dumps are
written under outputs/<experiment-name>/ (override with EXPERIMENT_ROOT).
cd eval
python -m pip install -e '.[api,data]'python -m zwz_eval init --directory local/my-eval --bench hr4k,hr8kEdit local/my-eval/model.json with your model/judge paths, identities, GPU settings
and output directory. Set downloaded annotation/image paths and input formats in
local/my-eval/datasets.json. Replace all example values before running.
Download and extract the benchmarks, then prepare them:
bash scripts/prepare_benchmarks.sh local/my-evalThis writes local/my-eval/ready.json without modifying the original downloads.
bash scripts/evaluate.sh local/my-evalThe command evaluates the configured benchmarks serially, managing model startup,
inference, required judging and service cleanup. To select a subset, append IDs
such as hr4k,hr8k.
MMStar, BabyVision, VStar, HR-4K, HR-8K, MME-RealWorld-en/cn, ZoomBench,
Perception-Bench, TreeBench, O3-Bench and CountQA require JSONL or a JSON array
with images, query and response fields:
{
"images": ["/path/to/images/image.png"],
"query": "<image>Original question and answer choices",
"response": "Reference answer"
}Use format: normalized for these benchmarks; keep original questions, answers,
image order and optional fields unchanged. Other benchmarks use their downloaded
formats. See eval/README.md for the benchmark ID table, download
instructions, data formats and optional API settings. AgentVista uses the same
workflow and additionally requires Internet access for search/visit tools.
This project builds upon:
- Qwen3.5 and Qwen3.6 for the base models.
- verl for the RL training framework.
For questions or collaborations, please contact:
- Lai Wei: waltonfuture@sjtu.edu.cn
- Yatai Li: zenith-knecht@sjtu.edu.cn
- Liangbo He: liangbo.hlb@antgroup.com
- Jun Lan: yelan.lj@antgroup.com
- Zhuosheng Zhang: zhangzs@sjtu.edu.cn
- Weiran Huang: weiran.huang@sjtu.edu.cn
If you encounter issues with dataset paths or evaluation setup, please share your feedback or contributions. Thank you for your support!
Apache License 2.0. See LICENSE.
@article{wei2026zoomingv2,
title={Zooming without Zooming v2: From Seeing Details to Connecting Evidence for Generalizable Multimodal Perception},
author={Wei, Lai and Li, Yatai and He, Liangbo and Lan, Jun and Gu, Renjie and Cao, Linhan and Lin, Junkai and Zhu, Huijia and Wang, Weiqiang and Kong, Linghe and Wang, Yue and Zhang, Zhuosheng and Huang, Weiran},
journal={arXiv preprint arXiv:2610.37037},
year={2026}
}