| XIFENG YAN |
home | publications | tutorials | dataset | software |
|
|
|||
|
|
Professor
Ph.D. (2006) |
Computer Science Department
Direction |
|
I am a professor at the University of California at Santa Barbara. The primary objective of my
research is to explore foundation
models in artificial intelligence, leveraging these models for knowledge
discovery, and developing cross-disciplinary applications in areas like
finance, healthcare, and science. We have extensively innovated on graph mining, graph
data management and pioneered
transformer-based time series
forecasting. I am also a co-inventor of
ADL/Mica, an
agent first approach to conversational AI assistants.
News (Nov 2025): An invited talk will be given at Nvidia, "Adaptive Inference in Pretrained LLMs," [slides]. This talk will cover our recent studies, "Adaptive Layer-skipping in Pre-trained LLMs," and "Directed Multi-Token Decoding" which illustrate and exploit the substantial underutilized computation in LLMs. Our approach enables the generation of multiple tokens using only a subset of Transformer layers. Unlike speculative decoding, our method introduces no additional parameters, no auxiliary routines, and requires no post-generation verification. News (Nov 2025): My another amazing student, Ross Koval, has been conducting impressive research on leveraging large language models for financial forecasting. His two recent papers -- "Context-Aware Language Models for Forecasting Market Impact from Sequences of Financial News" and "Multimodal Language Models with Modality-Specific Experts for Financial Forecasting from Interleaved Sequences of Text and Time Series" -- demonstrated significant progress in this area. Through market neutral portfolio simulation, his approach improved the Sharpe ratio from 1.14 to 1.51. This domain holds tremendous potential, both in methodology and application. News (Oct 2025): I am invited to give an oral spotlight talk (24 among 418 accepted papers), "Adaptive Inference in Pre-trained LLMs: Layer-skipping," at the Conference on Language Modeling (COLM) [slides]. The talk is about our recent paper, "Adaptive Layer-skipping in Pre-trained LLMs."[arxiv] [model/code/data] [poster]. A similar talk was given at the 2025 IEEE/RSJ Int. Conf. on Intelligent Robots and Systems (IROS 2025). News (Aug 2025): I was thrilled to give a talk, "Multimodal Language Models for Accelerating Scientific Discovery," at the 6th International Conference on Data-Driven Plasma Science [slides]. It is my first visit to Santa Fe. This trip is very special to me as reading legendary stories from Los Alamos and articles about complexity science at the Santa Fe Institute during my childhood helped shape my career interests. News (Aug 2025): Give a talk and poster at Berkeley Agent AI Summit: "Transforming Agent-based Chatbots with Declarative Programming" [slides][paper] News (June 2025): The training dataset of our multimodal language model Open-Qwen2VL released in April has been downloaded 90,000+ times on HuggingFace. It is a 2B-parameter multimodal language model pre-trained on 0.36% of the 1.4T multimodal tokens used in Qwen2-VL, but outperforms the closed-source Qwen2-VL-2B across various multimodal benchmarks. [arXiv] [project website] [code] [model] [data] News Archive |
||||||
| RESEARCH INTERESTS | ||||||
|
|
||||||
|
|
||||||
|
|
||||
|
Last Modified: July 10, 2022
