🚀 Excited to introduce EvoAgent! A generic method to automatically extend expert agents to multi-agent systems via the evolutionary algorithm!
Paper📄:
Website🌐:
Code💻:
[1/6]
🔆From Persona to Personalization: A Survey on Role-Playing Language Agents
🔍 Dive into our comprehensive survey of RPLA technologies, their applications, and the exciting potential for human-AI coexistence.
📖 Paper:
[1/3]
Excited to share our
#ACL2024
main paper "ANALOGYKB: Unlocking Analogical Reasoning of Language Models with A Million-scale Knowledge Base"!
📖 Paper:
🔗 Data: .
I'm pleased to share that, my Google Scholar profile has just crossed 100+ citations today🎉. Really thanks
@jiangjie_chen
, a senior mentor who provides me with a lot of guidance.
Exciting News!🎉 My submission to
#ACL2023
has triumphed with the 🏆Outstanding Paper Award🏆! Heartfelt gratitude to the meticulous reviewers for valuing our efforts.
Thanks to my mentor, Deqing Yang, & the dynamic team that brought this paper to life, especially
@jiangjie_chen
Apart from my team. I am really, really, really thankful to one reviewer, who gave a lot of meaningful suggestions.
What is even more valuable is that this reviewer showed me that there are such sincere and meticulous reviewers in this community
This reviewer is my role model👍
Excited to share our
#ACL2024
main paper "ANALOGYKB: Unlocking Analogical Reasoning of Language Models with A Million-scale Knowledge Base"!
📖 Paper:
🔗 Data: .
Thanks for sharing! 🥰
Let's try EvoAgent! A generic method via the evolutionary algorithm to automatically extend the specialized agent to multi-agent systems🥳
1/n EVOAGENT: Revolutionizing Agentic AI with Evolutionary Algorithms
In an era where large language models (LLMs) are pushing the boundaries of what's possible in AI, a team of researchers has taken a bold leap forward. They've dared to ask: What if we could create AI systems
With much frustration, I won't be able to attend the
#ACL2023
conference in Toronto due to visa issues possibly tied to the recent Canada-China tensions. 😔 To all the fellows who can make it, I envy you guys! Have a wonderful and successful conference.
#ACL2023NLP
Your Language Agents Already Know How to Achieve High-level Goals
Presents SelfGoal, a framework to enhance an LLM-based agent's capabilities to achieve high-level goals.
LLMs can perform well at basic tasks but struggle to achieve high-level goals without access to detailed
We categorize personas into three types: 1) Demographic Persona, which leverages statistical stereotypes; 2) Character Persona, focused on well-established figures; and 3) Individualized Persona, customized through ongoing user interactions for personalized services.
[2/3]
Moreover, thanks to the great benchmarks, i.e., Logic Grid Puzzle, Trivia Creative Writing, Codenames Collaborative created by
@zhenhailongW
, MMMU created by
@xiangyue96
, ScienceWorld created by
@peterjansen_ai
, and TravelPlanner created by
@jianxie_
and
@ysu_nlp
.
🚀 My favorite one got into EMNLP findings!
- We introduce an analogical structure abduction task to evaluate the analogical reasoning ability of LLMs.
- This task is grounded in cognitive psychology which is better aligned with human cognition.
Paper:
Moreover, EvoAgent can also extend interactive agents to multi-agent systems in solving complete scientific tasks in dynamic, open-world environments (e.g., ScienceWorld) and consistently improve the performance of LLMs.
[4/6]
Current LMs still struggle to achieve human-like performance in analogical reasoning tasks due to a lack of resources for model training. In this work, we propose ANALOGYKB, a million-scale analogy knowledge base derived from existing knowledge graphs.
[2/4]
For Real-World Scenarios, we also select TravelPlanner and prove that EvoAgent can generate specialized agents. Therefore, the generated travel plans are more aligned with user preferences and commonsense rules.
[5/6]
ANALOGYKB identifies two types of analogies:1)analogies of the same relations, which can be directly extracted from KGs;2)analogies of analogous relations, identified with a selection and filtering pipeline enabled by LLMs, with minor human efforts for data quality control.
[3/4]
We consider the existing agent frameworks as the initial individual and then apply a series of evolutionary operators (e.g., mutation, crossover, selection, etc.) to generate multiple agents with diverse agent settings.
[2/6]
🚀 Explore the AI showdown in our auction arena! 🤖 LLM agents are tested in a simulated auction, mastering strategy, risk, and resource management while providing clear strategy evaluations! 🛠️
📢Explore AI's prowess in our *Auction Arena*! LLM agents battle it out in a multi-item auction🧠💰
[0/7] Check out our new paper: *Put Your Money Where Your Mouth Is: Evaluating Strategic Planning and Execution of LLM Agents in an Auction Arena*.
Project:
We first test EvoAgent on three NLP tasks (Logic Grid Puzzle, Trivia Creative Writing, Codenames Collaborative) and MMMU.
The results show the EvoAgent can provide consistent improvements among each LLM, proving its strong generalization by using diverse generated agents.
[3/6]
Do you feel like your AI anime characters are always out-of-character?
How do we evaluate this?
I am thrilled to introduce our work: InCharacter- a novel perspective to evaluate the personality fidelity of RPAs with psychological scales
For application, we choose the debate scenario used in MetaGPT, which includes two debaters with different opinions, leading to dull content generation.
Instead of manually assigning new roles, EvoAgent can extend debate team to more agents with diverse settings.
[6/6]
We've delved into how humans follow goal-oriented scripts in their everyday actions & asked, 'What if AI could do the same, but with nuanced constraints (like baking a cake for diabetics)?' 💭🍰
Check out our
#ACL2023NLP
paper: 🧵
A more detailed introduction to EvoAgent is presented in the paper! Great joint work with
@jiangjie_chen
and
@SongKaitao
🎉
Many thanks to the great
@MSFTResearch
Asia team for the support!
@kimtaehyeon610
Out of curiosity, I carefully read both of your papers (btw, this idea is very interesting. Great work!). To me, even without the language aspect, I think the methods feel quite similar. But ARR can be rolled. If so, there’s a possibility that he submitted to ARR in August.
According to Structure Mapping Theory, analogical reasoning is based on identifying common relational structures between two systems. In this paper, we demonstrate that word analogies do not adequately reflect the analogical reasoning ability of LMs to align with human cognition.
The empirical evidence underlines the continued challenges faced by LLMs, including ChatGPT and GPT-4, in mastering this task, signifying the need for future exploration to enhance their abilities.
Thanks for sharing!
How do we automatically extend the specialized agent to multi-agent systems? 🤔
Try EvoAgent, a generic method via the evolutionary algorithm without any extra human designs! 💫
EvoAgent: A Generic Method to Automatically Extend Expert Agents to Multi-Agent Systems via the Evolutionary Algorithm
Quick read:
Paper:
Project:
@siyu_yuan_
@jiangjie_chen
and
@SongKaitao
My typical day as a Member of Technical Staff at OpenAI:
[9:00am] Wake up
[9:30am] Commute to Mission SF via Waymo. Grab avocado toast from Tartine
[9:45 am] Recite OpenAI charter. Pray to optimization Gods. Learn the Bitter Lesson
[10:00am] Meetings (Google Meet). Discuss how to
EASYTOOL can significantly reduce token consumption and greatly help LLM-based agents retrieve, select, and invoke tools, achieving better performance on tasks that interact with the external world.
3/4
Are Large Language Models Fair Summarizers?
We live in a world of value pluralism where several values can be equally correct and fundamental, and yet in conflict with each other.
Paper:
Code and Data:
We present a comprehensive overview of current methodologies for RPLAs, followed by the details for each persona type, covering data sourcing, agent construction and evaluation. Afterward, we discuss the fundamental risks, existing limitations and future prospects of RPLAs.
[3/3]
We analyze and explore the limitations of current tool utilization in LLM-based agents and first point out the deficiencies of tool documentation that prevent LLMs from using tools.
2/4
@CassielYM
Yes! As shown in our paper (Table 1 and Table 3), EvoAgent provides consistent improvements on both open-sourced LLMs (i.e., LLama2-13B-Chat and Mistral-7B) and close-sourced LLMs (i.e., GPT-4, GPT-3.5 and Gemini).
Can your GPT schedule TODOs *efficiently*?
Introducing TimeArena, a Time-Aware simulated textual environment for language agents to complete multiple tasks in the *shortest time*, which means simulating realistic temporal & resource constraints!
Website:
Evaluations on a series of datasets of two analogical reasoning tasks (analogy recognition and generation) demonstrate that ANALOGYKB successfully enables both smaller LMs and LLMs to gain better analogical reasoning capabilities.
[4/4]