|
I am a CS PhD student at
Northwestern.
I am fortunate to be advised by the wonderful
Manling Li.
Email /
Github /
Semantic Scholar /
Zhihu /
CV
|
|
|
|
The growth of foundation models, while extremely rapid, has heightened the need to address the challenges arising from their expanding scale. My research focuses on foundation models' autonomy (RAGEN, MINT benchmark), efficiency (DeepSeek-V2, Expert-Specialized Tuning, Chain-of-Experts), and long-context understanding (Long Video Haystack&T*, RETA-LLM). |
See full list on Google Scholar or Semantic Scholar (Why I Love Semantic Scholar, and You Might Too)
Semantic Scholar uses AI-powered tools to summarize papers, highlight key phrases, and rank research by influence. This helps you find important studies faster. Its Semantic Reader helps you understand papers with skimming highlights and citation cards . You can also see how papers connect with citation graphs. While Google Scholar is great for broad searches, Semantic Scholar is smarter for finding high-quality and impactful research!
|
Zihan Wang*, Kangrui Wang*, Qineng Wang*, Pingyue Zhang*, Linjie Li*, Zhengyuan Yang, Kefan Yu, Minh Nhat Nguyen, Monica Lam, Yiping Lu, Kyunghyun Cho, Jiajun Wu, Li Fei-Fei, Lijuan Wang, Yejin Choi, Manling Li Github Repository [Code] 
We introduce RAGEN to train LLM reasoning agents via RL in multi-turn, stochastic environments. RAGEN is formulated with MDP and optimized through Reasoning-Interaction Chain Optimization (RICO). RAGEN-0.5B is trained across three agentic tasks, showing intriguing reasoning patterns. |
|
Zihan Wang*, Rui Pan*, Lu Yin*, Manling Li*, Shiwei Liu* Github Repository [Blog] [Code] 
We propose Chain-of-Experts (CoE), enabling sequential communication between MoE experts by processing tokens on outputs from other intra-layer experts. CoE achieves 17.6-42% lower memory usage and reduces Math task validation loss from 1.20 to 1.12 with comparable compute. |
|
Jinhui Ye*, Zihan Wang*, Haosen Sun, Keshigeyan Chandrasegaran, Zane Durante, Cristobal Eyzaguirre, Yonatan Bisk, Juan Carlos Niebles, Ehsan Adeli, Li Fei-Fei, Jiajun Wu, Manling Li CVPR 2025 [Project Page] [Dataset] Paper & Code coming soon We introduce LongVideoHaystack, a 480-hour video temporal search dataset with 15,092 human-annotated instances, where SOTA scores 2.1% Temporal F1. Our temporal search framework T* boosts GPT-4o from 50.5% to 53.1% and LLaVA-OV from 56.5% to 62.4% on LongVideoBench XL. |
|
Zihan Wang, Deli Chen, Damai Dai, Runxin Xu, Zhuoshu Li, Yu Wu EMNLP 2024 [Paper] [Code] 
We harness the Specialized Power of Experts in MoE LLMs through ESFT. By fine-tuning Down to 5% Experts in a layer, near-full performance can be achieved. |
|
Xingyao Wang*, Zihan Wang*, Jiateng Liu, Yangyi Chen, Lifan Yuan, Hao Peng, Heng Ji ICLR 2024 [Paper] [Project Page] [Code] 
We introduce MINT, a benchmark for evaluating LLMs in Multi-turn Interactions with tools and language feedback. MINT reveals several limitations in existing RLHF and SIFT methods on multi-turn interaction. |
|
DeepSeek AI (157 authors including Zihan Wang) [Paper] [Code] 
DeepSeek-V2 is a strong MoE model with 23B activated parameters. It achieves stronger performance compared to DeepSeek 67B, saving 42.5% training costs and boosting generation by up to 5.76x. |
|
|
|
|
|
|
Website design from Jon Barron |