Department of Electronic and Computer Engineering - Seminar - InfiAgent: A Multi-Tool Agent for AI Operating Systems

10:30am - 12:00pm
ECE Conference room 2517 (via Lifts 25/26)

Following the launch of GPT4-Agent, GPT4 has demonstrated its flexibility in utilizing
tools like Advanced Data Analytics (ADA, previously known as code interpreter) and DALLE3,
although the details of GPT4-Agent have not been fully disclosed. Over the past years, we
have intensively studied the core functionalities of GPT4, progressively developing a system
comparable to GPT4-Agent, named InfiAgent. Initially, we replicated Codex and discovered
that while existing models such as CodeLlama, StarCoder, and WizardCoder excel in
programming capabilities, they fall short in handling FreeformQA problems for coding. To
address this, we created InfiCoder—the first open-source model capable of handling text-tocode,
code-to-code, and freeform code-related QA tasks simultaneously. Building on this,
we developed InfiCoder-Eval (FreeformQA benchmark), which includes 270 high-quality automated
test questions. Our findings indicate that even GPT4 has room for improvement
in this area (achieving a score rate of only 59.13%). Based on InfiCoder, we launched the
InfiAgent framework, focusing on the field of data analysis. This framework first defines the
problem framework and evaluation objectives for data analysis. Then, in line with the data
analysis scenarios, we developed a specialized Agent system based on the React format and
LLM, effectively addressing data analysis challenges. This system integrates an LLM with
programming capabilities and a sandbox environment for executing Python code, generating
solutions and corresponding code through multiple rounds of dialogues. It is the industry’s first
Agent framework closest to the capabilities of ADA. Additionally, we expanded the application
scenarios of InfiAgent, including a mathematical assistant and multimodal LLM (MLLM)
reasoning tools, achieving excellent results in math and multimodal reasoning. Particularly
in multimodal reasoning, we found that there is significant room for improvement in the current
GPT4V (achieving a score rate of only 74.44%). These achievements not only reveal the
tremendous potential of InfiAgent but also showcase our possible directions in surpassing the
capabilities of GPT4.

 

讲者/ 表演者:
Dr. Hongxia Yang
ByteDance - US

Dr. Hongxia Yang, PhD from Duke University, has published over 100 papers in top-tier
conferences and journals, and holds more than 50 patents in the USA and China. She was awarded
the prestigious Super AI Leader (SAIL Award) at the 2019World Artificial Intelligence Conference,
the Second-Class National Science and Technology Progress Award in 2020, the First-Class Science
and Technology Progress Award from the Electronics Society in 2021, named among the Top 50
Women in Tech by Forbes China in 2022, and received the First-Class Science and Technology
Progress Award from the Ministry of Education in 2022. Currently, she serves as the Head of
Large Models at ByteDance-US. Previously, she worked as a researcher at IBM’s global research
center Watson, Chief Data Scientist for Computational Advertising at Yahoo!, an AI scientist at
Alibaba’s DAMO Academy, and a part-time researcher at Zhejiang University’s Shanghai Advanced

语言
英文
适合对象
教职员
研究生
主办单位
电子及计算器工程学系
新增活动
请各校内团体将活动发布至大学活动日历。