Create README.md

README.md

@@ -0,0 +1,40 @@
+# NoteVQA: Benchmarking Vision-Language Models on Real-Life Questions from Human Communities
+
+<div align="left">
+
+[![arXiv](https://img.shields.io/badge/arXiv-NoteVQA-red)](https://arxiv.org/abs/2609.15695)
+[![HuggingFace](https://img.shields.io/badge/HuggingFace-NoteVQA-blue)](https://huggingface.co/datasets/AllSpark-Research/NoteVQA)
+
+*AllSpark Team*
+
+</div>
+
+---
+
+## Abstract
+
+Vision-language models (VLMs) increasingly power consumer-facing AI search, yet evaluating them on the diversity of everyday visual questions remains challenging. Existing benchmarks typically target predefined capabilities—such as multi-hop retrieval or long-form synthesis—whereas real users pose photo-grounded questions spanning a long tail of everyday scenarios. Notably, despite rapid advances in VLMs, users of Xiaohongshu, a mainstream Chinese image-sharing platform, continue to seek assistance from other humans for everyday visual questions. Motivated by this observation, we curate **NoteVQA**, a benchmark of 252 items spanning 12 topical categories and 7 user intents. Each item is accompanied by a concise reference distilled from expert community responses, together with a human-audited interleaved reference answer that combines textual explanations with supporting visual evidence.
+
+We evaluate both short-answer correctness and interleaved-answer quality. To support the latter, we introduce **AgenticInterleave**, a single-agent ReAct framework for retrieval-supported answer generation, along with **IVR-12**, a 12-dimensional rubric for assessing the content, presentation, and image quality of interleaved answers. Across nine frontier VLMs, the highest short-answer accuracy is 52.8%, while augmenting Qwen3.5-397B-A17B with agentic search improves accuracy by only 2.0%. For interleaved answers, the same model running AgenticInterleave scores 3.52 under IVR-12, compared with 4.65 for the human-audited references, with the largest gap observed in content quality. These findings underscore the challenges that everyday visual questions pose for current VLMs, both in answer accuracy and in the quality of visually grounded explanations.
+
+
+## 1. Overview
+
+Existing VLM benchmarks emphasize controlled, capability-oriented tasks. In contrast, NoteVQA is derived from *naturally occurring* information needs: real questions that users posted to a human community after existing AI tools failed to satisfy them. This design yields three distinctive properties:
+
+1. **Ecological validity.** Questions reflect authentic, photo-grounded information needs from everyday life rather than synthetic or templated constructions.
+2. **Long-tail coverage.** The benchmark spans 12 topical categories and 7 user intents, capturing the diversity of real user queries.
+3. **Interleaved ground truth.** Beyond short answers, each item provides a human-audited reference that interleaves textual explanation with supporting images, enabling evaluation of multimodal answer *presentation*, not merely correctness.
+
+## 2. Citation
+
+If you find NoteVQA useful in your research, please cite:
+
+```bibtex
+@article{jiang2026notevqa,
+  title   = {NoteVQA: Benchmarking VLMs on Real-Life Questions from Human Communities},
+  author  = {Jiang, Haonan and Zhan, Guojian and Xie, Jiancong and Wan, Shijun and Zhao, Dongjia and Chen, Cheng and Liu, Yahui and Hu, Yao and Mu, Chuan},
+  journal = {arXiv preprint arXiv:2609.15695},
+  year    = {2026}
+}
+```