Yuankai Li

I am a senior year undergraduate majoring in Artificial Intelligence in the School of Data Science, Fudan University.

I am privileged to be a research intern at PLUS lab at UCLA advised by Prof. Nanyun (Violet) Peng and Dr. Jia-Chen Gu. Prior to that, I was a research intern at Shanghai AI Laboratory under the guidance of Prof. Dequan Wang and Prof. Zhijie Deng.

My research interests lie broadly in machine learning and natural language processing. Currently, I have an intense interest in LLMs, with a specific focus on:

• Exploring and expanding the capability of LLMs to reason reliably via training, alignment and in-context learning

• Building helpful and reliable AI agents and exploring the use of LLMs in embodied AI

• Building intelligence systems that communicate through natural language and continuously improve through interaction

Email  /  CV  /  GitHub  / 

Research

BRIEF: Bridging Retrieval and Inference for Multi-hop Reasoning via Compression

We develope a lightweight, T5-based approach that performs query-aware multi-hop reasoning by compressing retrieved documents into highly dense textual summaries to integrate into in-context learning.

SCP: Scalable and Customizable Generation of Planning-specific Corner Cases in Autonomous Driving

We introduce a novel approach that leverages LLMs to generate scalable and customizable planning-specific corner cases for autonomous driving systems. It is data-agnostic and training-free, incorporating intention-driven context by explicit natural language commands.

Dissecting Dissonance: Benchmarking Large Multimodal Models Against Self-contradictory Instructions

We introduce the idea of self-contradictory instructions in Large Multimodal Models(LMMs), emphasized its potential harm, and sought to benchmark and address this problem.

Other Projects

These include coursework and side projects.

Exploration on Supervised-Finetuning on Small Language Models

I finetune GPT2-small over the MOSS-002 SFT dataset and conducted various capability tests with the BigBench dataset.

Simple Machine Learning Framework

I redevelope a simple machine learning framework in Python that implements a backpropagation algorithm, various neural network architectures (like MLP and CNN), manifold statistical learning models (like HMM and CRF), etc.

Implicit QR Algorithm for Eigen Value Problems

I implement the implicit QR algorithm proposed by John G. F. Francis(1962) in Matlab. The algorithm can find the eigenvalue of any given matrix in a reasonable amount of time.

Miscellanea

Things you can't find on my CV.

I enjoy playing Minesweeper. My average 3BV/s is 1.2 in Flag mode and 0.7 in NF mode. I'm still improving my skills.

I once learnt the basics of Sindarin from this.

I favour Star Trek DS9, VOY and TNG.

I am interested in speedrunning in most roguelike games.

Design and source code from Jon Barron's website