About Me

Yifan Xu is currently a Senior ML Research Engineer at Coinbase, where he investigates the potential of using AI and blockchain technology together to develop innovative solutions. He is also working on large language models (LLM) as part of his ongoing research. Yifan is passionate about advancing the field of AI and using it to tackle complex and diverse problems.

Before joining Coinbase, Yifan completed his Ph.D. in Cognitive Science at UC San Diego, where he also received his B.S. in Computer Science and Engineering. Under the supervision of Professor Zhuowen Tu, his research is centered around addressing the intricate complexities within several domains of artificial intelligence, such as natural language processing, computer vision, and reinforcement learning. Yifan has published his work in top conferences such as ICLR, CVPR, ICCV, and ACL.

---

News

• A New US Patent filing Related to LLM+RAG: "Dynamic Document Retrieval In A Retrieval-Augmented Generation System" - Filed with the US Patent Office under No. 18/949,932. It introduces a method dynamically optimizes the number of retrieved documents by employing gradient and clustering algorithms to analyze similarity distances between document embeddings
• Two New US Patent Filings Related to LLMs: (1) "Content Generation Using Enhanced Actor-Critic Models" - Filed with the US Patent Office under Application No. 18/646,500. It introduces a multi-agent framework that incorporates a critic agent to enhance the quality of text generation.
  (2) "Text-to-SQL Model Anchor Query Generation" - Filed with the US Patent Office under Application No. 18/670,720. It describes a semantic search framework designed to enhance multi-table search accuracy.
  
• ML and Blockchain Summit 2024 Highlights: As program chair of the ML and Blockchain Summit at Coinbase, I'm excited to announce our successful event with over 1500 attendees. Key moments included a fireside chat with Vitalik Buterin on AI and blockchain interplay, and talks from top speakers from Google, AWS, Dune Analytics, Stanford, and USC. Topics covered decentralized cloud for AI and zero-knowledge proofs in machine learning. Interested in speaking? Contact me.
• Introducing BLIVA: A simple Multimodal LLM for Better Handling of Text-Rich Visual Questions! We also explored some real-world applications, including YouTube thumbnail analysis! Code on GitHub.
• CoaT Model Becomes Essential: Our CoaT, a Vision Transformer model, hosted on Hugging Face has achieved 60K+ downloads every month. It has become one of the favored Transformers in various domains, including Kaggle competitions, industry, and academia, underscoring its growing prominence.

---

Publication

* indicates equal contribution

• PhD Thesis

• "Exploring Visual Perception with Transformers and World Model Representation", authored by Yifan Xu
  Read the thesis | View the slides

• LLM/VLM

• "Dynamic Document Retrieval In A Retrieval-Augmented Generation System", held by Coinbase and authored by Y Xu, Pending United States Patent.
• "Content Generation Using Enhanced Actor-Critic Models", held by Coinbase and authored by Y Xu, G Alperovich, V Mahadevan Pending United States Patent.
• "Text-to-SQL Model Anchor Query Generation", held by Coinbase and authored by Y Xu, R KB, I Rustandi, V Mahadevan Pending United States Patent.
• "BLIVA: A Simple Multimodal LLM for Better Handling of Text-Rich Visual Questions", W Hu*, Y Xu*, Y Li, W Li, Z Chen, Z Tu, AAAI 2024

• Blockchain

• "Detection of Malicious On-Chain Programs", held by Coinbase and authored by Y Xu, George Liu, I Rustandi, Y MA Pending United States Patent. Read more about it in the blog post: Detecting the Undetectable: Coinbase ERC-20 Scam Token Detection System

• Nature Language Processing

• "Convolutions and Self-Attention: Re-interpreting Relative Positions in Pre-trained Language Models", TA Chang, Y Xu, W Xu, Z Tu, ACL 2021
• "Rethinking exposure bias in language modeling", Y Xu*, K Zhang*, H Dong, Y Sun, W Zhao, Z Tu, arXiv preprint

• Reinforcement Learning

• "On the Feasibility of Cross-Task Transfer with Model-Based Reinforcement Learning", Y Xu*, N Hansen*, Z Wang, YC Chan, H Su, Z Tu, ICLR 2023
• "Neural Program Synthesis By Self-Learning", Y Xu*, L Dai*, U Singh, K Zhang, Z Tu, arXiv preprint

• Computer Vision

• "Co-scale conv-attentional image transformers", W Xu*, Y Xu*, T Chang, Z Tu, ICCV 2021 (Oral)
• "Line Segment Detection Using Transformers without Edges", Y Xu*, W Xu*, D Cheung, Z Tu, CVPR 2021 (Oral)
• "Pose Recognition with Cascade Transformers", K Li, S Wang, X Zhang, Y Xu, W Xu, Z Tu, CVPR 2021
• "Attentional Constellation Nets for Few-Shot Learning", W Xu*, Y Xu*, H Wang*, Z Tu, ICLR 2021
• "Guided variational autoencoder for disentanglement learning", Z Ding*, Y Xu*, W Xu, G Parmar, Y Yang, M Welling, Z Tu, CVPR 2020

---

Services

• Organizer

• I am serving as the program chair for the Machine Learning and Blockchain Summit at Coinbase in 2023 and 2024. This summit offers a platform for experts in both machine learning and blockchain technology to come together. We look forward to next year’s gathering and hope you can join us for a valuable exchange of ideas. If you are interested in joining as a speaker or panelist, please reach out to me.

• Reviewer

• I have dedicated my time to reviewing for top-tier ML/AI conferences such as CVPR, ICCV, ECCV, NeurIPS, and ICLR since 2021.

---

Tweets by yfxu_eth

---

About This Site

This website content is hosted on the InterPlanetary File System (IPFS), a decentralized network for storing and sharing files. It is backed up by FileCoin, a decentralized storage network that uses a blockchain-based system, through the use of Fleek, a platform for building and deploying websites on IPFS. This setup makes the website more secure and resilient, and enables it to be accessed by a wider audience.

My DNS (Centralized) URL: yfxu.com
My ENS (Decentralized) Address: yfxu.eth.limo

---