Zhiyu Zoey Chen
226
posts
@ZhiyuChen4
Dallas, TX
Joined May 2018
Yiyang is one of my favorite podcasters, highly recommend her podcast “独树不成林” in Chinese
Here is my interview with Christopher Nolan: are you a bard for a civilization at dusk?
00:00
Then NLP is low-dim-input-and-output, and by sheer luck we also have oceans of data
A nice read and very insightful! I think fundamentally vision is a high-dim-input problem so historically people have chosen low-dim "intermediate representations" (e.g., classes, segmentation masks) as target outputs to make everything tractable. On the other hand, control is
🚀In our new work on RAG efficient search, we propose a hierarchical process reward to optimize each reasoning step during training by on-the-fly detection of suboptimal searches. We boosted the accuracy and slashed the over-search rate from over 27% to just 2.3%! 📄 Paper:
Is your agentic RAG system searching too much or too little? 🤔 We tackle this problem in our new paper, "HiPRAG: Hierarchical Process Rewards for Efficient Agentic Retrieval Augmented Generation”. We introduce a novel training method that teaches search agents how to reason
Our new benchmark has been accepted to #EMNLP2025 main conference! A well curated dataset to test LLM agents to reproduce LM research papers. ⬇️
📢Excited to share our new paper on building agentic RAG with efficient search behaviors! Agentic RAGs often over- or under-search, hurting performance and reliability. We analyze these failure modes and propose a simple RL algorithm that rewards high-certainty search decisions,
