Bo Yin, Xiaobin Hu, Chengming Xu, Ruolin Shen, Mo Yang, Jiangning Zhang, Peng-Tao Jiang, Cheng Tan, Shuicheng Yan

arXiv GitHub

We introduce SPOT-E, a plug-and-play test-time visual adaptation framework that keeps the VLM frozen while optimizing question-conditioned visual spotlights through entropy shaping, improving evidence utilization without model retraining.

Bo Yin, Xiaobin Hu, Xingyu Zhou, Peng-Tao Jiang, Yue Liao, Junwei Zhu, Jiangning Zhang, Ying Tai, Chengjie Wang, Shuicheng Yan

arXiv GitHub

We present a novel fine-tuning approach for diffusion models, termed Frequency-Energy constrained Routing Adaptation (FeRA), which uses frequency-domain analysis to guide the fine-tuning process, enhancing both efficiency and effectiveness.

Xiaobin Hu, Enpu Zuo, Lanping Hu, Kaiwen Yang, Dianshu Liao, Tianyi Zhang, Bo Yin, Yinsi Zhou, Shidong Pan, Xiaoyu Sun

arXiv Project

We present VPD-100K, a large-scale fine-grained visual privacy dataset with 100,000 images, 33 classes, and over 190,000 object instances, together with a frequency-enhanced lightweight mechanism for robust privacy detection.

Xinlei Yu, Chengming Xu, Zhangquan Chen, Bo Yin, Cheng Yang, Yongbo He, Yihao Hu, Jiangning Zhang, Cheng Tan, Xiaobin Hu, Shuicheng Yan

arXiv GitHub

We introduce L²-VMAS, a model-agnostic framework that enables visual multi-agent collaboration through dual latent memories, improving scalability while reducing token-heavy text communication.

Bo Yin, Qi Li, Runpeng Yu, Xinchao Wang

arXiv GitHub

We propose an instance-level auditing method that infers, from a model’s behavior, whether it was trained on an original prompt or an LLM-refined version of that prompt within a mixed training corpus.

Bo Yin, Xiaobin Hu, Jiaqi Zhao, Shuicheng Yan

arXiv

FrameMorrow predicts compact prospective tokens to select historical frames according to future information needs, improving long-horizon video generation across diverse models.

Qi Li^*, Bo Yin^*, Weiqi Huang^*, Ruhao Liu^*, Bojun Zou, Runpeng Yu, Jingwen Ye, Weihao Yu, Xinchao Wang

^* Equal contribution.

arXiv GitHub

We survey VLA safety across threats, defenses, evaluation, and deployment, highlighting the unique risks introduced by embodied vision-language-action systems.

Xiaobin Hu, Haoyang He, Bo Yin, Yu He, Lei Xie, Jiangning Zhang, Yu-Gang Jiang, Shuicheng Yan

arXiv

We introduce MambaADv2, a multi-class unsupervised anomaly detection framework that combines a frozen encoder with a Mamba-inspired decoder using Duality-enhanced State Space modules to jointly model global, local, and frequency information with linear complexity.

Bo Yin, Qi Li, Xinchao Wang

arXiv GitHub

We introduce FATE, an on-policy self-evolving framework that converts verifier-scored failure trajectories into repair supervision and uses Pareto-Front Policy Optimization to improve trajectory-level safety while preserving utility.

Bo Yin, Xingyi Yang, Xinchao Wang

arXiv GitHub

We propose a novel and effective fine-tuning paradigm that introduces learnable activation functions to adapt large pre-trained models to downstream tasks while preserving their computational efficiency.

Xueming Yan, Bo Yin, Yaochu Jin

arXiv

We propose a novel multi-objective reinforcement learning framework that leverages latent temporal causal structures to enhance generalization and significantly outperform state-of-the-art baselines in complex dynamic environments.