Blame for README.md

## 📈 ZoomGainBench results

Scores are percentages. Each language's overall score is weighted by the number of questions in each category. Bold marks the best result in each column. [ZoomGainBench results in the paper (Table 2)](https://arxiv.org/html/2611.14076v1)

| Model | EN Attr. | EN Count | EN OCR | EN Overall | ZH Attr. | ZH Count | ZH OCR | ZH Overall | Overall |
| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: |
| **Closed-source Frontier VLMs** |  |  |  |  |  |  |  |  |  |
| GPT-5.5 | 52.75 | 57.45 | 60.40 | 56.51 | 52.12 | 63.83 | 60.96 | 57.06 | 56.79 |
| Gemini-3.6-Flash | 51.27 | 42.55 | **74.71** | 60.62 | 48.52 | 45.74 | **71.42** | 58.20 | 59.41 |
| Claude-Opus-5 | 48.09 | 63.83 | 71.86 | 59.89 | 48.73 | 62.77 | 71.29 | 59.84 | 59.86 |
| GPT-5.6-Sol | **56.57** | 62.77 | 67.93 | **62.08** | **54.03** | 63.83 | 65.84 | **60.07** | **61.08** |
| **Agentic VLMs** |  |  |  |  |  |  |  |  |  |
| DeepEyes-7B | 24.79 | 20.21 | 41.05 | 31.42 | 24.36 | 22.34 | 34.46 | 28.55 | 29.99 |
| Thyme-7B | 26.69 | 34.04 | 43.88 | 34.85 | 22.88 | 31.91 | 40.31 | 31.30 | 33.07 |
| DeepEyesV2-7B | 22.25 | 41.49 | 43.70 | 33.37 | 22.67 | 34.04 | 43.62 | 32.83 | 33.10 |
| PixelReasoner-7B | 27.12 | 45.74 | 44.83 | 36.56 | 25.42 | 23.40 | 40.13 | 31.62 | 34.09 |
| **Fine-grained Enhanced Single Forward VLMs** |  |  |  |  |  |  |  |  |  |
| ZwZ-4B | 30.08 | 39.36 | 29.68 | 30.78 | 33.47 | 50.00 | 34.45 | 35.45 | 33.12 |
| Vision-OPD-4B | 35.59 | 59.57 | 45.69 | 42.23 | 39.62 | 59.57 | 43.80 | 43.31 | 42.77 |
| Vision-OPD-9B | 46.40 | 55.32 | 47.18 | 47.58 | 47.67 | 60.64 | 37.97 | 44.68 | 46.13 |
| ZwZ-8B | 42.58 | 55.32 | 50.53 | 47.23 | 44.70 | 52.13 | 51.77 | 48.47 | 47.85 |
| **Open-source General VLMs** |  |  |  |  |  |  |  |  |  |
| MiMo-VL-7B-RL | 23.73 | 34.04 | 20.06 | 23.11 | 23.31 | 41.49 | 32.06 | 28.82 | 25.96 |
| GLM-4.6V (106B) | 24.36 | 38.30 | 31.57 | 28.80 | 22.46 | 36.17 | 28.51 | 26.37 | 27.59 |
| Qwen3-VL-4B | 21.82 | 43.62 | 31.14 | 27.92 | 22.88 | 41.49 | 33.63 | 29.30 | 28.61 |
| Qwen3-VL-8B | 15.89 | 41.49 | 39.17 | 28.40 | 19.92 | 30.85 | 41.64 | 30.37 | 29.38 |
| Qwen3.5-4B | 21.40 | 51.06 | 34.40 | 29.83 | 22.67 | 46.81 | 34.87 | 30.23 | 30.03 |
| Qwen3.5-9B | 21.82 | 41.49 | 37.79 | 30.60 | 22.67 | 47.87 | 42.52 | 33.66 | 32.13 |
| Qwen3.5-35B-A3B | 28.18 | 44.68 | 46.10 | 37.51 | 27.33 | 51.06 | 47.95 | 38.51 | 38.01 |
| Qwen3.6-35B-A3B | 26.91 | 54.26 | 50.21 | 39.59 | 26.69 | 52.13 | 46.17 | 37.54 | 38.57 |
| Qwen3-VL-235B-A22B | 29.45 | 52.13 | 48.01 | 39.64 | 30.93 | 50.00 | 50.90 | 41.39 | 40.51 |
| Qwen3.5-27B | 33.05 | 51.06 | 45.92 | 40.33 | 34.75 | 58.51 | 48.67 | 43.02 | 41.68 |
| Qwen3.6-27B | 33.47 | 56.38 | 47.47 | 41.70 | 37.71 | **70.21** | 45.33 | 44.07 | 42.89 |
| Qwen3.8-27B | 40.25 | 54.26 | 45.49 | 43.84 | 44.92 | 62.77 | 43.05 | 45.78 | 44.81 |
| Qwen3.5-397B-A17B | 33.47 | 65.96 | 51.95 | 44.55 | 38.14 | 62.77 | 50.41 | 45.78 | 45.16 |
| GLM-5.3-Flash (320B) | 44.07 | 54.26 | 43.77 | 44.90 | 46.61 | 55.32 | 43.74 | 46.18 | 45.54 |
| Kimi-K3 (2.8T) | 49.15 | 64.89 | 45.48 | 49.04 | 46.82 | 64.89 | 46.22 | 48.26 | 48.65 |
| Qwen3.8-Flash-Next (125B) | 47.46 | **67.02** | 56.97 | 53.42 | 52.97 | 67.02 | 58.45 | 56.67 | 55.05 |

## 🖼️ Benchmark comparison

![ZoomGainBench compared with existing fine-grained perception benchmarks](docs/figures/zoomgain-benchmark-comparison.png)

[Comparison of benchmark design in the paper (Table 1)](https://arxiv.org/html/2611.14076v1)

# 🔍 ZoomGain

<p align="center"><strong>Mining Global-to-Zoom Perception Gaps for Fine-Grained Visual Supervision</strong></p>

<p align="center">
  <a href="https://github.com/inclusionAI/ZoomGain">💻 GitHub</a> &nbsp;·&nbsp;
  <a href="https://arxiv.org/abs/2611.14076">📄 Paper</a> &nbsp;·&nbsp;
  <a href="https://huggingface.co/datasets/inclusionAI/ZoomGainBench">📊 ZoomGainBench</a> &nbsp;·&nbsp;
  <a href="https://huggingface.co/datasets/inclusionAI/ZoomGain-RL-Data">🤗 RL Data</a> &nbsp;·&nbsp;
  <a href="https://huggingface.co/inclusionAI/ZoomGain-4B">🧠 ZoomGain-4B</a> &nbsp;·&nbsp;
  <a href="https://huggingface.co/inclusionAI/ZoomGain-9B">🧠 ZoomGain-9B</a>
</p>

<p align="center">
  <img src="docs/figures/zoomgain-overview.png" width="920" alt="ZoomGain uses zoomed-in details to build questions about the full image, producing a challenging benchmark and training data that improves fine-grained perception." />
</p>

**ZoomGain** builds questions about a full image from visual details that become clear when you zoom in. We use this global-to-zoom perception gap to construct supervision for vision-language models. In the paper, the best of 28 models scored **61.08** on ZoomGainBench, and ZoomGain-9B reached superior performance on multiple fine-grained perception and general perception benchmarks. [Paper](https://arxiv.org/abs/2611.14076)

## 📜 Updates

- **`2026.11.11`** 🌟 We released the [arXiv paper](https://arxiv.org/abs/2611.14076).
- **`2026.11.11`** 🌟 We released the evaluation benchmark and RL training dataset. [ZoomGainBench](https://huggingface.co/datasets/inclusionAI/ZoomGainBench) · [ZoomGain RL Data](https://huggingface.co/datasets/inclusionAI/ZoomGain-RL-Data)
- **`2026.11.11`** 🌟 We released the model checkpoints. [ZoomGain-4B](https://huggingface.co/inclusionAI/ZoomGain-4B) · [ZoomGain-9B](https://huggingface.co/inclusionAI/ZoomGain-9B)

## 🔬 How ZoomGain Works

<p align="center">
  <img src="docs/figures/zoomgain-pipeline.png" width="920" alt="Three stages of ZoomGain data construction: anchor grounding, global-to-zoom evidence mining, and answer verification with multiple-choice distractors." />
</p>

1. **Ground the target.** Multiple models independently locate a target described in the full scene. Their agreement identifies a region to crop.
2. **Mine evidence.** Compare captions of the same region at both scales, then turn details revealed by the crop into questions about the full image.
3. **Verify questions and answers.** Cross-check answers on the crop, filter ambiguous questions and questions answerable without the image, and build plausible distractors for multiple-choice items.

## 🛠️ Setup

Clone the repository and set an asset directory. The commands below assume you run them from the repository root.

```bash
git clone https://github.com/inclusionAI/ZoomGain.git
cd ZoomGain

export ASSET_DIR="$PWD/zoomgain-assets"
mkdir -p "$ASSET_DIR"
python -m pip install -U huggingface_hub
```

Run `hf auth login` if the Hugging Face repositories require authenticated access. [Hugging Face CLI guide](https://huggingface.co/docs/huggingface_hub/guides/cli)



### 🧰 Install inference and evaluation dependencies

```bash
python -m pip install -r eval_ZoomGainBench/requirements.txt \
  'vllm==0.19.1' 'transformers==5.7.0'
```


## 📊 Evaluate on ZoomGainBench

### 📦 Download the benchmark

ZoomGainBench contains **1,000 questions over 941 images** covering attribute recognition, counting, and OCR. Each annotation includes English and Chinese question text, a reference answer, scoring rules, and an evidence bounding box in image pixel coordinates.

```bash
BENCH_DIR="$ASSET_DIR/data/ZoomGainBench"
hf download inclusionAI/ZoomGainBench annotations.jsonl images.zip \
  --repo-type dataset --local-dir "$BENCH_DIR"
unzip -q "$BENCH_DIR/images.zip" -d "$BENCH_DIR"
```

Keep `images/` next to `annotations.jsonl`. The evaluator loads the full image from each annotation's relative `image_path` and uses `evidence_bbox` to validate the evidence location.

### 🚀 Run evaluation

Install the dependencies from the model section and run the benchmark from the repository root. Replace the example `MODEL_PATH` with the path to a local checkpoint directory for any compatible model.

```bash
MODEL_PATH="local/path/to/Qwen3.5-4B" \
ANNOTATIONS="$BENCH_DIR/annotations.jsonl" \
OUTPUT_DIR="outputs/qwen3.5-4b-bench" \
GPUS=0 PORT=18811 LANGUAGE=all \
bash eval_ZoomGainBench/eval.sh
```

- `LANGUAGE=all` evaluates both languages for **2,000 tasks**. Use `LANGUAGE=en` for English or `LANGUAGE=cn` for Chinese.
- `GPUS=0,1` with `TP=1` runs two independent replicas. Set `TP=2` to use both GPUs for one replica.
- Append `--dry-run` to validate the model and dataset or `--limit 10` for a short GPU run. Evaluation defaults to temperature `0` with Qwen3.5 thinking disabled.

Read `report.md` for the evaluation scores, `summary.json` for aggregate metrics, and `results.json` for per-question results. `responses.jsonl` contains the raw model responses.

## 🎯 Train on ZoomGain RL Data

### 📦 Download the training data

The release contains **9,000 training records** and **7,713 source images**. The Parquet stores prompts, relative image paths, reference answers, and metadata used by the reward function.

```bash
RL_DIR="$ASSET_DIR/data/ZoomGain-RL-Data"
hf download inclusionAI/ZoomGain-RL-Data train.parquet images.zip \
  --repo-type dataset --local-dir "$RL_DIR"
unzip -q "$RL_DIR/images.zip" -d "$RL_DIR"
```

The training launcher resolves image paths relative to `train.parquet` and gives EasyR1 a temporary Parquet with absolute paths. Keep the extracted `images/` directory next to `train.parquet`.

```text
ZoomGain-RL-Data/
├── train.parquet
└── images/
    ├── image_000001.jpg
    └── ...
```

### ⚙️ Install the training environment

Use a Python 3.10 environment with a CUDA toolkit and a C++ compiler for any CUDA extensions that need to build locally. The example below uses the CUDA 12.8 PyTorch wheels. [PyTorch wheel index](https://pytorch.org/get-started/previous-versions/#v2100)

```bash
python -m pip install torch==2.10.0 torchvision==0.25.0 \
  --index-url https://download.pytorch.org/whl/cu128
python -m pip install packaging psutil ninja
python -m pip install --no-build-isolation flash-attn==2.8.3 causal-conv1d==1.6.2.post1
python -m pip install -r requirements.txt
```

The launch scripts load the bundled EasyR1 code directly. Point `BASE_MODEL` at the original Qwen3.5 checkpoint to start RL training from the same model family used in the paper. [Qwen3.5-4B](https://huggingface.co/Qwen/Qwen3.5-4B) · [Qwen3.5-9B](https://huggingface.co/Qwen/Qwen3.5-9B)

### 🧪 Validate and run a short training job

```bash
hf download Qwen/Qwen3.5-4B --local-dir "$ASSET_DIR/models/Qwen3.5-4B"

python zoomgain/prepare_data.py --input "$RL_DIR/train.parquet"

BASE_MODEL="$ASSET_DIR/models/Qwen3.5-4B" \
TRAIN_PARQUET="$RL_DIR/train.parquet" \
OUTPUT_DIR="$PWD/outputs/train-smoke-4b" \
CUDA_VISIBLE_DEVICES=0,1 ROLLOUT_N=4 MAX_RESPONSE_LENGTH=512 ROLLOUT_AGENT_WORKERS=2 \
bash scripts/smoke_train.sh
```

The smoke test runs one GRPO step, including rollout generation, reward calculation, and a checkpoint save.

### 🏋️ Launch a full training run

```bash
BASE_MODEL="$ASSET_DIR/models/Qwen3.5-4B" \
TRAIN_PARQUET="$RL_DIR/train.parquet" \
OUTPUT_DIR="outputs/train-zoomgain-4b" \
CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 \
bash scripts/train_zoomgain_4b.sh
```

For 9B, download `Qwen/Qwen3.5-9B` to a local directory, update `BASE_MODEL` and `OUTPUT_DIR`, and run `scripts/train_zoomgain_9b.sh`. The full launchers default to one epoch, a batch size of 64, and eight rollouts per question. The output directory contains training logs, rollouts, and checkpoints. To evaluate a saved model, set `MODEL_PATH` to `<output-dir>/checkpoints/global_step_<N>/actor/huggingface`.



## 📄 License

The code is released under the Apache License, Version 2.0. The bundled training framework retains its upstream license and notices. See the Hugging Face cards for the terms of each dataset and checkpoint. [LICENSE](https://github.com/inclusionAI/ZoomGain/blob/main/LICENSE) · [NOTICE](https://github.com/inclusionAI/ZoomGain/blob/main/NOTICE)

## ⭐ Citation
If you find this repository helpful to your research, welcome to cite our paper:
```bibtex
@article{shen2026zoomgain,
  title   = {ZoomGain: Mining Global-to-Zoom Perception Gaps for Fine-Grained Visual Supervision},
  author  = {Shen, Haozhan and He, Liangbo and Lan, Jun and Zhu, Huijia and Wang, Weiqiang and Zhao, Tiancheng and Yin, Jianwei},
  journal = {arXiv preprint arXiv:2611.14076},
  year    = {2026}
}
```