YANAMI

A Large-Scale Multimodal Dataset and Benchmark for Multiturn Stateful Interaction

Shengyu Wang1*† Tengyue Nong2* Junming Liu3* Haosong Peng1* Siyu Li4* Ruiqi Lei5 Fushuo Huo6 Yuhao Pan1 Yunfeng Fan7 Ding Wang8 Nan Cheng9 Haozhao Wang10 Wenchao Xu1‡
1 The Hong Kong University of Science and Technology 2 Independent Researcher 3 The University of Melbourne 4 The Chinese University of Hong Kong, Shenzhen 5 Imperial College London 6 Southeast University 7 The Hong Kong Polytechnic University 8 Shanghai Artificial Intelligence Laboratory 9 Xidian University 10 Huazhong University of Science and Technology

* Equal contribution (co-first authors)† Project leader‡ Corresponding author

Paper Dataset Benchmark
402,639 dialogues 2,047,236 images 2,643,291 turns 400-dialogue human-verified benchmark
YANAMI overview: previous multi-turn work performs sequential, independently solvable edits, while YANAMI captures multimodal dialogue with stateful interaction across four scenarios, a 400K-dialogue training set and a human-verified benchmark.
Figure 1. YANAMI is the first large-scale multimodal dialogue dataset and benchmark for stateful interaction — 400K dialogues, 2M+ images, 2.6M turns, plus a 400-dialogue human-verified benchmark.
The research question

Can a model keep the whole conversation in the picture?

Users establish constraints, refer to earlier images and revise decisions. A successful response must recover the currently valid dialogue state and render it faithfully. Many existing multi-turn tasks can be solved one instruction at a time; YANAMI makes the accumulated history essential.

YANAMI — Your Assistant Needs to Assimilate Multiturn Interaction — brings together large-scale training data and a human-verified benchmark across multi-image fusion, image editing, world modeling and visual reasoning.

Contributions

What YANAMI provides

Training data, a controlled benchmark and evaluation of both understanding and visual realization.

YANAMI-400K

402,639 dialogues, 2M+ images and 2.6M turns built around six state-management capabilities.

YANAMI-Bench

400 human-verified dialogues, 1,600 understanding questions and 8,467 atomic scoring points.

Two-axis evaluation

Dialogue State Understanding (four capability QAs) is scored jointly with Visual Realization (State Consistency + Visual Quality), with judge–human correlations above 0.80.

Real benchmark dialogues

Four scenarios. A conversation to remember.

Explore five real dialogues in each scenario, with the original messages and images shown in conversation.

Dialogue state understanding

Capability Q&A

Four original benchmark questions and reference answers for this dialogue, shown separately from the conversation.

Complete original dialogues and reference / ground-truth images from YANAMI-Bench. More cases on Hugging Face ↗
Overall results

Stateful multi-turn interaction remains challenging

Fine-tuning BAGEL on YANAMI-400K lifts its benchmark score from 65.02 to 78.25, while preserving general visual understanding.

+13.23
YANAMI-Bench points · 65.02 → 78.25
+10.28%
relative GEdit-EN gain · 6.52 → 7.19
+2.0 pp
GenEval · 0.73 → 0.75

YANAMI-Bench overall

16 systems evaluated on dialogue state understanding and visual realization. Higher is better.

YANAMI
Citation

BibTeX

If you use YANAMI in your research, please cite the manuscript below.

@misc{wang2026yanami,
  title     = {{YANAMI}: A Large-Scale Multimodal Dataset and Benchmark
               for Multiturn Stateful Interaction},
  author    = {Wang, Shengyu and
               Nong, Tengyue and
               Liu, Junming and
               Peng, Haosong and
               Li, Siyu and
               Lei, Ruiqi and
               Huo, Fushuo and
               Pan, Yuhao and
               Fan, Yunfeng and
               Wang, Ding and
               Cheng, Nan and
               Wang, Haozhao and
               Xu, Wenchao},
  year      = {2026},
  note      = {Research manuscript}
}