Zhongsheng Wang

Time:

• 15:00-17:00 Beijing Time
• Dec 19, 2024 (Thursday)

Venue:

518, Research Building 4

Abstract:

Large language models (LLMs) are widely used in data generation and reasoning. This talk covers the basic concepts, advantages, development history, and pre-training and fine-tuning process of LLMs, and focuses on some of the speaker’s work: the combined application of LLMs in high-quality self-supervised data generation, LLM agents combined with reinforcement learning to generate epic long text story novels, and challenges and solutions in reasoning tasks, including multi-step logical reasoning using LLM-generated code. In addition, we propose a potential idea: to improve the performance of LLMs in multi-step deductive logical reasoning tasks by building a world model and combining it with a model checking method (LLM+X).

Speaker Bio:

Zhongsheng Wang is the PhD Candidate at UoA, focusing on Trustworthy AI Agents based on MRAG. He received his undergraduate degree in Computer Science and Technology from the Southwest University in 2023 and master degree in Data Science at the University of Auckland in 2024. His research interests include natural language processing, and he’s currently focusing on research on LLM-Based Agents and the practical application of LLMs. He’s also a member of the LIU AI LAB, which is dedicated to cutting-edge research in artificial intelligence.

Euclid in around  300 BC proved that the sequence of prime numbers is infinite. This sequence starts 2,3,5,7,11, 13, …; the largest currently known prime number  is  obtained by raising 2 to the power of 82589933 and subtracting 1. Each number is a unique product of prime numbers; so the prime numbers  can be seen as the building blocks for all natural numbers. 

The first part of the talk gives an overview of prime numbers, including a fairly recent result of Green and Tao on arithmetic progressions,  and close calls to the Goldbach conjecture due to Wang, Helfgott, and others. 

The second part focusses on computation: how to recognise via an efficient (i.e., polynomial time) algorithm whether a number is prime, and how to use a hypothetical quantum computer to obtain the prime factorisation of a number in polynomial time. The relevance to cryptography will be discussed as well.

Speaker Bio

PhD 1992, Heidelberg

University of Chicago, 1995-2001

University of Auckland, 2002-

2010 ICM speaker in special session

2020 Humboldt research award

该论文研究了著名的聚类点删除问题（CVD问题）的算法与计算复杂性，得到该问题精确求解中，以点删除集大小k为参数的最佳运行时间上界，改进了Tsur于2021年给出的结果。本文给出的时间运行界和历史上该问题的运行时间如表1中所示。

CVD问题是图算法中一个著名的NP难问题，该问题问是否能通过删除图中不超过k个顶点使得剩下图中每一个连通块都是一个完全图。由于CVD问题良好地刻画了聚类的性质，该问题在机器学习和计算生物学中有着广泛的应用。本论文首先针对低度图上的CVD问题进行了分析，在度数不超过4的图上先设计了一个快速的算法；然后在4度图的结果基础上，使用自动生成搜索树（Automated Generation of Searching Trees）的技术对一般图进行了深入分析，最终改进该问题当前最好的算法。

田康毅同学从大一开始进入算法与逻辑团队学习，在肖鸣宇教授的指导下从事参数算法与核心化算法相关方向的研究工作，已经研究获得多项科研成果，形成学术论文两篇。

主讲人：

傅育熙，1992年获英国曼彻斯特大学计算机博士学位，1994年起在上海交通大学任职，历任计算机系主任，软件学院院长。是国家杰出青年基金获得者、上海市优秀学科带头人。研究领域为理论计算机科学，内容涉及程序理论、并发理论、等价性验证、可达性理论、交互理论等。学术兼职有：上海高校软件理论研究中心主任、国务院学位委员会第六届学科评议组成员 (2010-2014)、上海市计算机学会理事长 (2015-2018）、教育部计算机类专业教学指导委员会副主任 （2013-2017，2018-2022）。是Mathematical Structuresin Computer Science期刊的编委。傅育熙讲授的 《计算复杂性理论》课程获得“2019年度高校计算机专业优秀教师奖励计划”。

课程包含两部分：一、随机计算与去随机，主要内容包括：随机算法、概率图灵机与BPP、通用哈希函数族、随机游走、扩张图及其显式构造、Reigold定理。二、交互证明系统，主要内容包括：私币交互证明系统、公币交互证明系统、两类交互证明系统的等价性、IP=PSPACE、多证明者交互证明系统。
参考教材：《计算复杂性理论》（傅育熙著、清华大学出版社，2023年）。

上课时间：

17-17周，星期一第3-4节 第5-6节

星期二第3-4节第5-6节

星期三第3-4节 第5-6节

星期四第3-4节 第5-6节

星期五第1-4节

上课地点：立人楼A108