Publications

2026

PMI
Free Lunch for Stabilizing Rectified Flow Inversion
Chenru Wang, Beier Zhu, Chi Zhang
ICLR 2026 (The Fourteenth International Conference on Learning Representations)
We develop two plug-and-play velocity correction methods to improve inversion stability and editing fidelity in flow-based generative models. Achieved state-of-the-art results on PIE-Bench.
IMS3
IMS3: Breaking Distributional Aggregation in Diffusion-Based Dataset Distillation
Chenru Wang, Yunyi Chen, Zijun Yang, Joey Tianyi Zhou, Chi Zhang
CVPR 2026 (IEEE/CVF Conference on Computer Vision and Pattern Recognition)
We develop an inversion-based fine-tuning method for dataset distillation and dataset generation, along with a cluster-based group sampling method for distilled data sampling.
ManimAgent cross-task memory overview
ManimAgent: Self-Evolving Multimodal Agents for Visual Education
Wenjia Jiang, Zongyuan Cai, Yuanhang Shao, Chenru Wang, Boyan Han, Zhixue Song, Keyu Chen, Shengwei An, Xu Yang, Zhou Yang
EMNLP 2026 (The 2026 Conference on Empirical Methods in Natural Language Processing)
ManimAgent carries reflection experience across tasks through a self-grown, dual-channel Episodic Memory Bank. It retains successful visual-education patterns and validated failure lessons without weight updates or human-seeded examples.
ScoutSampler: Learning to Find Causally Critical Frames for Long Video Understanding
Wenjia Jiang*, Chenru Wang*, Yiwei Wang, Yanming Yang, Boyan Han, Keyu Chen, Shengwei An, Zhou Yang, Chi Zhang
NeurIPS 2026 Workshop (Grounded and Faithful Vision-Language Models for Real-World Deployment)

2025

DDUM
DDUM: Deformable Dilated U-structure Module for Coronary Stenosis Detection
Chenru Wang, Zirui Chen, Muyao Li, Haoran Yin, Saijie Zhou, Jingliang Zhang, Xueying Zeng, Qing Zhang
Medical Engineering & Physics, 2025
We apply deep learning techniques to assist physicians in diagnosing coronary artery disease. We propose the Deformable Dilatable U-structure Module to specialize generic networks for coronary stenosis detection and improve their generalization ability. We also construct the 302 dataset for evaluation in real-world clinical scenarios.

Preprints

StepX-OPD
StepX-OPD: Rethinking Multi-Teacher On-Policy Distillation in Diffusion Models
Chenru Wang, Xuanyang Zhang, Yucheng Han, Wenbo Nie, Wei Cheng, Xianfang Zeng, Gang Yu, Chi Zhang
Preprint, 2026
We train a single few-step text-to-image diffusion model by distilling from multiple reward-specialized teachers on the student's own rollouts, keeping the few-step trajectory faithful via direct velocity regression on low-noise batches.
Rethinking Backdoors in Text-to-Image Generation Models: When, Why, and How
Chenru Wang, Neelesh Samptur, Sheng-Yen Chou, Shengfang Zhai, Pin-Yu Chen, Xiangyu Zhang, Shengwei An
Preprint, 2026
To Ask or Not to Ask: Backdooring the Learned Consultation Policy of Collaborative LLM Agents
Chenru Wang, Tong Li, Wenjia Jiang, Boyan Han, Xiaoxi Wei, Chengzhong Wang, Zhou Yang, Yan Chen, Shengwei An
Preprint, 2026
MGPO: Manifold-Guided Diffusion Alignment for Task-Aware Dataset Distillation
Yunyi Chen*, Chenru Wang*, Xinyi Ye, Zexin Zheng, Chi Zhang
Preprint, 2026
Immune Escape in Model Immunization
Tong Li*, Chenru Wang*, Chris Thomas, Shengwei An
Preprint, 2026
GTS: Gap-Targeted Synthesis for Budget-Efficient Synthetic Data Scaling
Hantian Zhang*, Chenru Wang*, Xinyi Ye, Xiaoxi Wei, Chi Zhang
Preprint, 2026