README.md

---
configs:
- config_name: default
  data_files:
  - split: test
    path: metadata/eval_metadata.jsonl
task_categories:
- text-to-image
license: apache-2.0
---

<div align="center">

# Search Beyond What Can Be Taught

### Evolving the Knowledge Boundary in Agentic Visual Generation

Haozhe Wang<sup>1</sup> · Weijia Feng<sup>3</sup> · Jinpeng Yu<sup>3</sup> · Che Liu<sup>4</sup> · Ping Nie<sup>2</sup> · Fangzhen Lin<sup>1</sup> · Jiaming Liu<sup>3</sup>✉ · Ruihua Huang<sup>3</sup> · Jimmy Lin<sup>2</sup> · Wenhu Chen<sup>2</sup> · Cong Wei<sup>2</sup>✉

<sup>1</sup> HKUST · <sup>2</sup> University of Waterloo · <sup>3</sup> Qwen Applications · <sup>4</sup> Imperial College London

✉ Corresponding authors: Jiaming Liu, Cong Wei

[📄 arXiv](https://arxiv.org/abs/2607.05382) · [🌐 Project Page](https://haozheh3.github.io/SearchGen/) · [💻 GitHub](https://github.com/HaozheH3/SearchGen)

[🤗 SearchGen-20K](https://huggingface.co/datasets/JasperHaozhe/SearchGen-20K) · [🤗 SearchGen-Corpus-1M](https://huggingface.co/datasets/JasperHaozhe/SearchGen-Corpus-1M) · [🤗 AgentGen-Bench](https://huggingface.co/datasets/JasperHaozhe/AgentGen-Bench)

**Image generators fabricate what they don't know. _This one looks it up first—and knows when not to._**

</div>

## AgentGen-Bench

This dataset contains the standalone AgentGen benchmark prompts and reference metadata.

## Dataset Statistics

| Statistic | Count |
| --- | ---: |
| Benchmark prompts | 751 |
| Visual reference images | 1,099 |

## Official Evaluation Subsets

> [!IMPORTANT]
> AgentGen-Bench includes three official compact-evaluation labels:
> **`testmini`**, **`testmini_easy`**, and **`testmini_hard`**.

| Subset label | Rows | Intended track |
| --- | ---: | --- |
| `testmini` | 200 | Compact evaluation |
| `testmini_easy` | 100 | Compact easy track |
| `testmini_hard` | 100 | Compact hard track |

Each row stores its memberships in the `subset` array. The labels are
independent membership tags, so do not infer one label from another. Rows that
do not belong to these subsets have `"subset": []`.

The published **Test-mini** result uses exactly the 200 rows carrying the
`testmini` label. It is a fixed benchmark slice, not a fresh or random sample
from the 751-row Full benchmark.

Example with Hugging Face Datasets:

```python
from datasets import load_dataset

benchmark = load_dataset("JasperHaozhe/AgentGen-Bench", split="test")
testmini = benchmark.filter(lambda row: "testmini" in row["subset"])
testmini_easy = benchmark.filter(lambda row: "testmini_easy" in row["subset"])
testmini_hard = benchmark.filter(lambda row: "testmini_hard" in row["subset"])

assert len(testmini) == 200
```

## Benchmark Results

Scores are on a 0–100 scale. Each table reports Overall9 for the same 18
generators. Coverage is `scored prompts / slice prompts`, with missing
evaluator scores excluded from aggregation.

### Full benchmark (751 prompts)

| Rank | Model | Coverage | Overall9 |
|---:|---|---:|---:|
| 1 | GPT-Image-2 | 718/751 | 76.0 |
| 2 | Grok-Imagine-2.0-Low | 715/751 | 75.2 |
| 3 | Qwen-Image-3-Max | 738/751 | 68.4 |
| 4 | Grok-Imagine | 717/751 | 67.8 |
| 5 | Nano Banana Pro | 738/751 | 65.0 |
| 6 | Qwen-Image-2-Pro | 742/751 | 59.4 |
| 7 | Qwen-Image-2 | 741/751 | 56.4 |
| 8 | SeedDream-4.5 | 746/751 | 56.2 |
| 9 | SeedDream-4.0 | 750/751 | 52.5 |
| 10 | Nano-Banana | 740/751 | 49.7 |
| 11 | SenseNova-U1 | 750/751 | 35.7 |
| 12 | Qwen-Image | 751/751 | 34.1 |
| 13 | Mage-Flow | 751/751 | 32.8 |
| 14 | Flux.2-Klein-9B | 750/751 | 31.1 |
| 15 | Flux.2-Klein-4B | 751/751 | 27.7 |
| 16 | Bagel | 751/751 | 25.8 |
| 17 | OmniGen2 | 750/751 | 24.0 |
| 18 | Show-o2 | 751/751 | 19.8 |

### Test-mini (200 prompts)

| Rank | Model | Coverage | Overall9 |
|---:|---|---:|---:|
| 1 | GPT-Image-2 | 200/200 | 76.3 |
| 2 | Grok-Imagine-2.0-Low | 196/200 | 73.9 |
| 3 | Qwen-Image-3-Max | 195/200 | 68.2 |
| 4 | Grok-Imagine | 196/200 | 67.7 |
| 5 | Nano Banana Pro | 200/200 | 64.7 |
| 6 | Qwen-Image-2-Pro | 197/200 | 60.3 |
| 7 | Qwen-Image-2 | 197/200 | 56.9 |
| 8 | SeedDream-4.5 | 199/200 | 56.6 |
| 9 | SeedDream-4.0 | 200/200 | 51.8 |
| 10 | Nano-Banana | 200/200 | 48.4 |
| 11 | SenseNova-U1 | 200/200 | 35.4 |
| 12 | Qwen-Image | 200/200 | 32.7 |
| 13 | Mage-Flow | 200/200 | 31.4 |
| 14 | Flux.2-Klein-9B | 200/200 | 30.5 |
| 15 | Flux.2-Klein-4B | 200/200 | 27.1 |
| 16 | Bagel | 200/200 | 24.8 |
| 17 | OmniGen2 | 199/200 | 22.2 |
| 18 | Show-o2 | 200/200 | 19.1 |

<details>
<summary><strong>Mini-easy and Mini-hard leaderboards (100 prompts each)</strong></summary>

`testmini_easy` and `testmini_hard` are independent official label slices;
they are not necessarily a partition of the 200-prompt `testmini` slice.

### Mini-easy

| Rank | Model | Coverage | Overall9 |
|---:|---|---:|---:|
| 1 | GPT-Image-2 | 100/100 | 79.7 |
| 2 | Grok-Imagine-2.0-Low | 98/100 | 79.5 |
| 3 | Qwen-Image-3-Max | 99/100 | 74.9 |
| 4 | Grok-Imagine | 98/100 | 73.4 |
| 5 | Qwen-Image-2-Pro | 99/100 | 69.3 |
| 6 | Nano Banana Pro | 100/100 | 66.5 |
| 7 | SeedDream-4.0 | 100/100 | 64.3 |
| 8 | SeedDream-4.5 | 99/100 | 63.9 |
| 9 | Qwen-Image-2 | 99/100 | 63.5 |
| 10 | Nano-Banana | 100/100 | 54.4 |
| 11 | SenseNova-U1 | 100/100 | 40.6 |
| 12 | Qwen-Image | 100/100 | 38.7 |
| 13 | Mage-Flow | 100/100 | 37.7 |
| 14 | Flux.2-Klein-9B | 100/100 | 34.8 |
| 15 | Flux.2-Klein-4B | 100/100 | 32.6 |
| 16 | Bagel | 100/100 | 29.3 |
| 17 | OmniGen2 | 100/100 | 26.5 |
| 18 | Show-o2 | 100/100 | 20.2 |

### Mini-hard

| Rank | Model | Coverage | Overall9 |
|---:|---|---:|---:|
| 1 | Grok-Imagine-2.0-Low | 96/100 | 71.9 |
| 2 | GPT-Image-2 | 100/100 | 71.3 |
| 3 | Qwen-Image-3-Max | 99/100 | 64.3 |
| 4 | Grok-Imagine | 94/100 | 61.8 |
| 5 | Nano Banana Pro | 100/100 | 60.8 |
| 6 | Qwen-Image-2-Pro | 100/100 | 50.6 |
| 7 | Qwen-Image-2 | 100/100 | 50.0 |
| 8 | SeedDream-4.5 | 100/100 | 49.0 |
| 9 | Nano-Banana | 100/100 | 44.4 |
| 10 | SeedDream-4.0 | 100/100 | 38.9 |
| 11 | SenseNova-U1 | 100/100 | 29.7 |
| 12 | Qwen-Image | 100/100 | 29.5 |
| 13 | Mage-Flow | 100/100 | 28.3 |
| 14 | Flux.2-Klein-9B | 100/100 | 27.3 |
| 15 | Bagel | 100/100 | 23.0 |
| 16 | Flux.2-Klein-4B | 100/100 | 21.9 |
| 17 | OmniGen2 | 100/100 | 21.9 |
| 18 | Show-o2 | 100/100 | 15.8 |

</details>

## Purpose

Use this section to evaluate image generation systems on prompts with explicit visual references, text knowledge slots, verification checklists, and rubrics. It is separate from the 20K training metadata and is intended for evaluation.

## Main Files

- `eval_metadata.jsonl`: benchmark prompt metadata.
- `benchmark_manifest.json`: benchmark counts, schema fields, and reference-image policy.
- `eval_query_coverage.jsonl`: coverage map from benchmark reference slots to `searchgen-corpus-1m` queries.

## Schema

Each `eval_metadata.jsonl` row contains:

- `bench_id`
- `subset`
- `user_prompt`
- `language`
- `generation_task_type`
- `failure_modes`
- `domains`
- `verification_checklist`
- `evaluation_rubric`
- `text_knowledge_slots`
- `visual_reference_slots`

`visual_reference_slots[].relative_path` points to the benchmark-local reference image path. It is relative to `data_release/searchgen-bench/`. After repairing source eval-dataset reference paths, all 1,099 benchmark reference-image targets are materialized. The benchmark-only copy audit is `../../validation_logs/payload_copy_benchmark_reference_after_fix.tsv`; the current missing-image worklist `../../validation_logs/missing_benchmark_reference_images.tsv` is empty.

Expected final image layout:

```text
data_release/searchgen-bench/reference_images/{bench_id}/ref_00.jpg
```

## Cross-References

`eval_query_coverage.jsonl` links benchmark reference slots to `../searchgen-corpus-1m/database/search.sqlite` through `query_id` when a matching search corpus query exists.

The training dataset is separate and documented in `../searchgen-20k/README.md`.

Minimal load check:

```bash
python3 - <<'PY'
import json
rows=[json.loads(x) for x in open('data_release/searchgen-bench/eval_metadata.jsonl')]
print(len(rows), rows[0]['bench_id'])
PY
```

## Rights and takedowns

Rights questions and takedown requests: `jasper.whz@outlook.com`.

## Citation

If you find this dataset useful, please cite:

```bibtex
@article{wang2026searchgen,
  title   = {Search Beyond What Can Be Taught: Evolving the Knowledge
             Boundary in Agentic Visual Generation},
  author  = {Wang, Haozhe and Feng, Weijia and Yu, Jinpeng and Liu, Che and
             Nie, Ping and Lin, Fangzhen and Liu, Jiaming and Huang, Ruihua and
             Lin, Jimmy and Chen, Wenhu and Wei, Cong},
  journal = {arXiv preprint arXiv:2607.05382},
  year    = {2026},
  url     = {https://arxiv.org/abs/2607.05382}
}
```