1

Analytical Correction for Subsampling Bias in Drifting Models

We show that the minibatch centroid in drifting models carries an O(1/n) bias from softmax self-normalization, and propose Analytical Bias Correction (ABC), a two-line closed-form fix that reduces the bias to O(1/n²).

Jiaru Zhang, Zeyun Deng, Juanwu Lu, Ziran Wang, Ruqi Zhang

Analytical Correction for Subsampling Bias in Drifting Models

MLS-Bench: A Holistic and Rigorous Assessment of AI Systems on Building Better AI

We introduce MLS-Bench, a benchmark of 140 tasks across 12 ML domains that evaluates whether AI systems can invent generalizable and scalable ML methods, and find that current agents remain far from reliably surpassing human-designed methods.

Bohan Lyu, Yucheng Yang, Siqiao Huang, Jiaru Zhang, Qixin Xu, Xinghan Li, Xinyang Han, Yicheng Zhang, Huaqing Zhang, Runhan Huang, Kaicheng Yang, Zitao Chen, Wentao Guo, Junlin Yang, Xinyue Ai, Wenhao Chai, Yadi Cao, Ziran Yang, Kun Wang, Dapeng Jiang, Huan-ang Gao, Shange Tang, Chengshuai Shi, Simon S. Du, Max Simchowitz, Jiantao Jiao, Dawn Song, Chi Jin

MLS-Bench: A Holistic and Rigorous Assessment of AI Systems on Building Better AI

Efficient and Explainable End-to-End Autonomous Driving via Masked Vision-Language-Action Diffusion

MVLAD-AD combines masked vision-language-action diffusion with discrete action tokenization, geometry-aware embeddings, and action-priority decoding for efficient, precise, and explainable end-to-end autonomous driving.

Jiaru Zhang, Manav Gagvani, Can Cui, Juntong Peng, Ruqi Zhang, Ziran Wang

ViLaD: A Large Vision Language Diffusion Framework for End-to-End Autonomous Driving

Can Cui, Jiaru Zhang, Yupeng Zhou, Juntong Peng, Sung-Yeon Park, Zichong Yang, Prashanth Sankaranarayanan, Ruqi Zhang, Ziran Wang

Exploring Diffusion Models' Corruption Stage in Few-Shot Fine-tuning and Mitigating with Bayesian Neural Networks

Xiaoyu Wu, Jiaru Zhang, Yang Hua, Bohan Lyu, Hao Wang, Tao Song, Haibing Guan

Exploring Diffusion Models' Corruption Stage in Few-Shot Fine-tuning and Mitigating with Bayesian Neural Networks

One-Step Diffusion Samplers via Self-Distillation and Deterministic Flow

Self-Distilled One-Step Diffusion Samplers achieve fast one/few-step sampling and stable evidence estimation through state-space and volume consistency with a deterministic-flow importance weight.

Pascal Jutras-Dubé, Jiaru Zhang, Ziran Wang, Ruqi Zhang

Stealthy Backdoor Attack in Federated Learning via Adaptive Layer-wise Gradient Alignment

Qingqian Yang, Peishen Yan, Xiaoyu Wu, Jiaru Zhang, Tao Song, Yang Hua, Hao Wang, Liangliang Wang, Haibing Guan

Leveraging Model Guidance to Extract Training Data from Personalized Diffusion Models

Diffusion Models (DMs) have evolved into advanced image generation tools, especially for few-shot fine-tuning where a pretrained DM is …

Xiaoyu Wu, Jiaru Zhang, Steven Wu

Learning Identifiable Structures Helps Avoid Bias in DNN-based Supervised Causal Learning

In this paper, we propose a novel DNN-based method for supervised causal learning that addresses systematic biases in existing methods, with a newly designed pairwise encoder serving as the core architecture.

Jiaru Zhang, Rui Ding, Qiang Fu, Bojun Huang, Zizhen Deng, Yang Hua, Haibing Guan, Shi Han, Dongmei Zhang

Learning Identifiable Structures Helps Avoid Bias in DNN-based Supervised Causal Learning

CGI-DM: Digital Copyright Authentication for Diffusion Models via Contrasting Gradient Inversion

in this paper, we are the first to explore and propose to utilize adversarial examples for DMs to protect human-created artworks. Specifically, we first build a theoretical framework to define and evaluate the adversarial examples for DMs. Then, based on this framework, we design a novel algorithm, named AdvDM, which exploits a Monte-Carlo estimation of adversarial examples for DMs by optimizing upon different latent variables sampled from the reverse process of DMs.

Xiaoyu Wu, Yang Hua, Chumeng Liang, Jiaru Zhang, Hao Wang, Tao Song, Haibing Guan

CGI-DM: Digital Copyright Authentication for Diffusion Models via Contrasting Gradient Inversion