Multimodal large models — training, RLVR post-training, and trustworthy evaluation. Advised by Prof. Yiqiang Chen.
“When a benchmark score goes up, did the model actually acquire the ability the benchmark claims to measure?”
01News
CRAFT — The Diagnosis Is Already in the Notes — accepted to Findings of EMNLP 2026 (first author).
KnowMeBenchV2 accepted to the EMNLP 2026 main conference (co-author).
Received the ICT Outstanding Master’s Student Award.
Invention patent granted: a training method for large models in ophthalmology.
VersaFusion presented at AAAI 2025 (first author).
02Publications
Does RLVR teach vision-language models to see, or to better use what they already saw? A train × test visual-visibility study, plus FlipTrack — a counterfactual benchmark of paired images whose answers must flip, so image-blind strategies fail by construction. Code and benchmark will be released with the paper.
Diagnosis-redacted clinical records still recover the diagnosis at 0.944 macro-AUROC. CRAFT is a portable audit protocol that reports each channel’s shortcut floor, so fusion gains are read against what the text alone already gives away.
@inproceedings{ye2026diagnosis,
title = {The Diagnosis Is Already in the Notes: A Channel-Wise
Audit of Diagnostic Shortcuts in Clinical-Text Benchmarks},
author = {Ye, Haocun and Jiang, Xinlong and Hu, Yuanyuan and Chen, Shubai
and Wu, Tingyu and Zheng, Zhenkun and Yi, Caijiao and Zhang, Teng
and Wang, Liuxin and Zhao, Tong and Xu, Heping and Dai, Weiwei
and Chen, Yiqiang},
booktitle = {Findings of the Association for Computational Linguistics:
EMNLP 2026},
year = {2026},
note = {Code: \url{https://github.com/yhc98002-bit/The-Diagnosis-Is-Already-in-the-Notes}}
}
An evaluation-only benchmark for evidence-grounded, person-centric long-video understanding.
@inproceedings{knowmebenchv2_2026,
title = {KnowMeBenchV2: Evidence-Grounded Person-Centric
Long-Video Understanding},
author = {Wu, Tingyu and Cao, Guangyu and Lan, Qizhen and Chen, Zhisheng
and Weng, Ziyan and Zhu, Bingkun and Su, Miao and Li, Chenglong
and Xie, Zhengwei and Wang, Huacan and Hu, Sen and Ye, Haocun
and An, Nan and Liu, Zaoqu and Chen, Ronghao},
booktitle = {Proceedings of the 2026 Conference on Empirical Methods
in Natural Language Processing (EMNLP 2026)},
year = {2026}
}
A tuning-free dual-branch diffusion framework — proximal negative-prompt inversion, KV memory injection and classifier guidance — for fine-grained editing that leaves everything else alone. AAAI 39(9), 9445–9453.
@inproceedings{ye2025versafusion,
title = {VersaFusion: A Versatile Diffusion-Based Framework for
Fine-Grained Image Editing and Enhancement},
author = {Ye, Haocun and Jiang, Xinlong and Gao, Chenlong
and Wang, Bingyu and Huang, Wuliang and Chen, Yiqiang},
booktitle = {Proceedings of the AAAI Conference on Artificial Intelligence},
volume = {39},
number = {9},
pages = {9445--9453},
year = {2025},
doi = {10.1609/aaai.v39i9.33023},
url = {https://ojs.aaai.org/index.php/AAAI/article/view/33023}
}
Patent · Granted A Training Method for Large Models in Ophthalmology — invention patent behind the ophthalmology LLM training pipeline.
03Projects
Does RLVR teach vision-language models to see, or to better use what they already saw? A causal train × test visual-visibility study on Qwen2.5-VL, plus FlipTrack — a counterfactual benchmark of paired images whose answers must flip, so any image-blind strategy fails one member of every pair by construction.
A channel-wise audit protocol for clinical-text benchmarks: diagnosis-redacted records still recover the diagnosis at 0.944 macro-AUROC, so fusion gains must be reported alongside the text branch’s shortcut floors. Now being engineered into a leakage-audit module for our lab’s data-governance platform.
From 11 open medical QA datasets and 212 OCR’d textbooks to 9.0M ophthalmic SFT samples, then multi-node full-parameter SFT of Qwen2.5-14B and Qwen2-VL-7B with DeepSpeed ZeRO-3 — report-generation CIDEr rose from 0.0004 to 0.557. Industry-sponsored; method granted an invention patent.
Tuning-free dual-branch diffusion editing (AAAI 2025) — 5.23 s preparation versus 63.56 s for DragDiffusion — deployed on the MADA platform to simulate how a crop will look at later growth stages from a single photo, while keeping the plant’s identity and the scene unchanged.
04About
I’m Haocun Ye, a PhD student in artificial intelligence at the Institute of Computing Technology, Chinese Academy of Sciences, advised by Prof. Yiqiang Chen.
I work on multimodal large models — how they are trained, how they are post-trained with reinforcement learning, and how they should be evaluated. The question that drives most of my work is simple to state and hard to answer: when a benchmark score goes up, did the model actually acquire the ability the benchmark claims to measure?
Away from the lab I like experimenting in the kitchen, and I’m happiest on days off spent with family, travelling with friends, or around a board-game table — ideally with a dog nearby. In quieter hours I play single-player games (GTA, The Witcher 3, Sekiro). I know I have my share of flaws, and I try to keep improving — both in how I work with others and in myself.
05Contact
Open to research collaboration and conversations about multimodal models, post-training, and evaluation. Email is the fastest way to reach me.
Institutional: yehaocun23s@ict.ac.cn