HumanScale: Egocentric Human Video Can Outperform Real-Robot Data for Embodied Pretraining

Juncheng Ma*, Jianxin Bi*, Yufan Deng, Xuanran Zhai, Kewei Zhang, Ye Huang, Bo Liang, Shukai Gong, Jiankai Tu, Xiaotian Tang, Jiaxin Li, Kaiqi Chen, Duomin Wang, Yuqi Wang, Bingyi Kang, Eric Huang, Zhiyang Dou, Zhen Dong, Enze Xie, Wojciech Matusik, Tat-Seng Chua, Daquan Zhou

(* stands for equal contribution)

Paper | Code

With a carefully designed filtering and labeling pipeline, we show that egocentric human video is a scalable pretraining source for embodied foundation models that surpasses real-robot data, especially in out-of-distribution generalization.

StableVLA: Towards Robust Vision-Language-Action Models without Extra Data

Yiyang Fu, Chubin Zhang, Shukai Gong, Yufan Deng, Kaiwei Sun, Qiyang Min, Qibin Hou, Yansong Tang, Jianan Wang3†, Daquan Zhou1††

Paper | Code

We propose a lightweight adapter module for VLA, IB-Adapter, which selectively filters potential noise from visual inputs. Without requiring any extra data or augmentation strategies, IB-Adapter consistently improves the robostness of VLA, while adding fewer than 10M parameters.

TPP-SD: Accelerating Transformer Point Process Sampling with Speculative Decoding

Shukai Gong*, Yiyang Fu*, Fengyuan Ran*, Quyu Kong, Feng Zhou†

(* stands for equal contribution)

Paper | Code

By identifying the structural similarities between thinning algorithms for TPPs and speculative decoding for language models, we develop TPP-SD, which maintains the identical output distribution as autoregressive sampling while achieving 2-6× speedup on both synthetic and real datasets.

USPTO-LLM: A Large Language Model-Assisted Information-enriched Chemical Reaction Dataset

Shen Yuan*, Shukai Gong*, Hongteng Xu†

(* stands for equal contribution)

Paper | Dataset

we construct an information-enriched chemical reaction dataset called USPTO-LLM, which comprises over 247K chemical reactions extracted from the patent documents of USPTO encompassing abundant information on reaction conditions. Experiments show that USPTO-LLM helps pre-train the existing retrosynthesis methods and the condition information in the dataset helps improve the model performance.

Research Intern, Bytedance Seed, Beijing, 2026.04 - present

Research on physical video generation and embodied AI.

Research Intern, Pixverse, Beijing, 2025.07 - 2026.01

Research on the acceleration and long-video adaption of visual-audio joint generation.