范文棟-從工具到自主化構建更強大的Agent系統.pdf

《范文棟-從工具到自主化構建更強大的Agent系統.pdf》由會員分享，可在線閱讀，更多相關《范文棟-從工具到自主化構建更強大的Agent系統.pdf（38頁珍藏版）》請在三個皮匠報告上搜索。

1、ML-SummitML-SummitML-SummitML-SummitML-SummitML-SummitML-SummitML-SummitML-SummitML-SummitML-SummitML-SummitML-SummitML-SummitML-SummitFinding the Scaling Laws of AgentsWendong FanCore Member of CAMEL-AITech Lead of Eigent AICAMEL-AI.org從具到主化，構建更強的Agent系統范棟 CAMEL-AIML-SummitML-SummitML-SummitAgent f

2、rom 1986 Agents are mindless processes Agent by itself can only do some simple things Joining these agents in societies leads to true intelligenceWhat magical trick makes us intelligent?The trick is that there is no trick.The power of intelligence stems from our vast diversity,not from any single,pe

3、rfect principle.Marvin Minsky,The Society of Mind,p.308ML-SummitML-SummitML-SummitThe agent-environment interaction in RLThe learner and decision-maker is called the AgentAgent interacts with an environment through actionsEnvironment presents new states and rewardsAgentsAgent in Reinforcement Learni

4、ngML-SummitML-SummitML-SummitGame AgentAutomated VehiclesGo GameAgentsAgent in Reinforcement LearningML-SummitML-SummitML-SummitML-SummitML-SummitML-SummitLilian Weng:https:/lilianweng.github.io/posts/2023-06-23-agent/Key differences:Language as InputLanguage as OutputState and Action are expressed

5、as natural languageGeneralizability.Language Models as AgentsML-SummitML-SummitML-SummitLanguage Models as AgentsKey Features:-Memory:Manages chat history and context window-Tools:Supports both internal and external function calls-Step Loop:Handle task require multiple request with one stepML-Summit

6、ML-SummitML-Summitclass TaskSpecifyAgent(ChatAgent):An agent that specifies a given task prompt by prompting the user toprovide more details class TaskPlannerAgent(ChatAgent):An agent that helps divide a task into subtasks based on the inputtask prompt class TaskCreationAgent(ChatAgent):An agent tha

7、t helps create new tasks based on the objectiveand last completed task.Compared to:obj:TaskPlannerAgent,its still a task planner,but it has more context informationlike last task and incomplete task list.class TaskPrioritizationAgent(ChatAgent):An agent that helps re-prioritize the task list andretu

8、rns numbered prioritized list.from camel.agents import(TaskCreationAgent,TaskPlannerAgent,TaskPrioritizationAgent,TaskSpecifyAgent)PlanningAgentsTask DecompositionSelf-ReflectionLanguage Model as AgentML-SummitML-SummitML-Summit#Set the role name and the taskcritic_role=a picky critic#Create the met

9、a_dict and the role_tuplemeta_dict=dict(critic_role=critic_role,criteria=Help better accomplish the task.)#Generate the system messagesys_msg=sys_msg_gen().from_dict(meta_dict=meta_dict,role_tuple=(critic_role,RoleType.CRITIC)critic_agent=CriticAgent(system_message=sys_msg,verbose=True)print(critic_

10、agent.system_message.content)You are a a picky critic who teams up with a user_role and a assistant_role to solve a task:task.Your job is to select an option from their proposals and provides your explanations.Your selection criteria are Help better accomplish the task.You always have to choose an o

11、ption from the proposals.PlanningTask DecompositionSelf-ReflectionAgentsLanguage Model as AgentML-SummitML-SummitML-SummitML-SummitML-SummitML-SummitThe new era of Multi-agent system with LLM agents?#Parameters-#Agents?Data-Environments?Training-Memory/Interaction?The next scaling laws?Agent scaling

12、 law hypothesis:ML-SummitML-SummitML-SummitMar 21,2023Mar 21,2023CAMELApr.03,2023Apr.03,2023Baby AGIMar.30,2023Mar.30,2023AutoGPTJul.16,2023Jul.16,2023ChatDevAug.01,2023Aug.01,2023MetaGPTNov.15,2023Nov.15,2023CrewAIThe FirstMulti-Agent FrameworkLLM based Agent SystemsML-SummitML-SummitML-Summit Idea

13、 Role assignment Task agents Chat agentsRole-Playing FrameworkCAMEL(NeurIPS 2023):https:/arxiv.org/abs/2303.17760CAMELML-SummitML-SummitML-SummitCAMEL(NeurIPS 2023):https:/arxiv.org/abs/2303.17760 Idea Role assignment Task agents Chat agentsCAMELML-SummitML-SummitML-SummitCAMEL(NeurIPS 2023):https:/

14、arxiv.org/abs/2303.17760WorkforceHierarchical ArchitectureTask Planning and DecompositionTask Channel CommunicationRobust Failure HandlingCAMELML-SummitML-SummitML-SummitAgent Evaluation Results CAMEL agents are better one agent 70%on 200 tasks GPT-4 evaluation aligns with Human evaluationMore Agent

15、s Between than one?CAMEL(NeurIPS 2023):https:/arxiv.org/abs/2303.17760ML-SummitML-SummitML-SummitOWL:Optimized Workforce LearningML-SummitML-SummitML-Summit 58.18 average score on GAIA benchmark Ranking#1 among open-source frameworks!OWL:Optimized Workforce LearningML-SummitML-SummitML-SummitCRAB:Cr

16、oss-environment Agent Benchmark for Multimodal Language Model Agents(NeurIPS 2024 OWA workshop):https:/arxiv.org/pdf/2407.01511CRAB:Cross-environment Agent Benchmark For Multimodal Language Model AgentsML-SummitML-SummitML-SummitCrab:Cross-environment Agent Benchmark forMultimodal Language Model Age

17、ntsTask:Open slack,navigate to multi-modal-benchmark channel,summarize the last two messages,and then send a message to the summary to the first contact on the phoneCRAB Agents and EnvironmentsML-SummitML-SummitML-SummitOASIS:Open Agent Social Interaction Simulations with One Million Agents(NeurIPS

18、2024 OWA workshop):https:/arxiv.org/abs/2411.11581 Up to 1 million agents Replicate social science experiments Explore dynamic of agent societyOASIS:Simulate Social Media with1 million AgentsML-SummitML-SummitML-SummitML-SummitML-SummitML-SummitAgents with GraphRAGML-SummitML-SummitML-SummitRole-Pla

19、ying Scraper for Report&Knowledge Graph GenerationCAMEL ImplementationAutonomous Task SolvingKnowledge GenerationScalable SolutionAgents with GraphRAGML-SummitML-SummitML-SummitRole-Playing Scraper for Report&Knowledge Graph GenerationAutonomous Task SolvingKnowledge GenerationScalable SolutionCAMEL

20、 ImplementationAgents with GraphRAGML-SummitML-SummitML-SummitRole-Playing Scraper for Report&Knowledge Graph GenerationAutonomous Task SolvingKnowledge GenerationScalable SolutionCAMEL ImplementationAgents with GraphRAGML-SummitML-SummitML-Summithttps:/docs.camel-ai.org/cookbooks/sft_data_generatio

21、n_and_unsloth_finetuning_tinyllama.htmlAgentic Data GenerationML-SummitML-SummitML-Summithttps:/www.camel-ai.org/data_explorerAI Society:50 assistant roles,50 user roles and 10 tasks Code:20 programming languages,50 domains,and 50 tasksAgentic Data GenerationML-SummitML-SummitML-SummitCAMEL(NeurIPS

22、2023):https:/arxiv.org/abs/2303.17760Finetune LLMs with CAMEL DatasetsGenerate data from GPT3.5/4Finetune Llama modelsAgentic Data GenerationML-SummitML-SummitML-Summithttps:/huggingface.co/models?dataset=dataset:camel-ai%2Fmath&sort=downloadsFinetune LLMs withCAMEL Datasets180+models on Huggingface

23、MosaicML MPT-30B-ChatHuggingFaceM4/idefics2-8bAgentic Data GenerationML-SummitML-SummitML-SummitFinding the scaling law of agents:#Parameters-#Agents?-CAMEL,OASIS Data-Environments?-CRAB,LOONG,OWL Training-Evolution?-DataGen,GraphRAGThe Next Scaling Laws?ML-SummitML-SummitML-SummitCAMEL is a open source multi-agent system frameworkML-SummitML-SummitML-SummitML-SummitML-SummitML-Summit