【重点报告】清华大学-大模型工具学习（中英文）-2024.pptx

工具学习;背景;工具和情报;工具和情报;工具和情报;工具学习的分类;工具学习的分类;框架;框架;意图理解;意图理解;工具理解;工具理解;规划与推理;规划与推理;规划与推理;规划与推理;规划与推理;培训策略;WebGPT;WebCPM;WebCPM;WebCPM;WebCPM;WebCPM;WebCPM;Webshop;Toolformer;工具创建;工具创建;工具创建;工具创建;工具创建;应用程序;ChatGPT插件;开源解决方案;开源解决方案;开源解决方案;开源解决方案;开源解决方案;开源解决方案;开源解决方案;开源解决方案;开源解决方案;开源解决方案;Summary;工具学习论文列表;谢谢;秦禹嘉THUNLP;THUNLP;Toolsareextensionsofhumancapabilitiesdesignedtoenhanceproductivity,efficiency,andproblem-solving

Throughouthistory,humanshavebeentheprimaryagentsintheinventionandmanipulationoftools

Question:canartificialintelligencebeascapableashumansintooluse?;Theanswerisyeswithfoundationmodels

Strongsemanticunderstanding

Extensiveworldknowledge

Powerfulreasoningandplanningcapabilities…;ToolsandIntelligence;Tool-augmentedlearning

Augmentfoundationmodelswiththeexecutionresultsfromtools

Toolsareviewedascomplementaryresourcesthataidinthegenerationofhigh-qualityoutputs;CategorizationofToolLearning;THUNLP;Framework;9;Scalingupthemodelsizeandthediversityofinstruction-tuningdatasets

Enhancementofgeneralizationcapability

Challenges

UnderstandingVagueInstructions:vaguenessandambiguityintheuserquery

TheoreticallyInfiniteInstructionSpace:infiniteexpressionandpersonalizedinstructions;ToolUnderstanding;ToolUnderstanding;IntrospectiveReasoning

Generateastaticplanwithoutinteractingwiththeenvironment

ExtrospectiveReasoning

Generateadynamicplanconsideringthechangeofenvironmentandfeedbacks;IntrospectiveReasoning

Ifpromptedappropriately,PLMscaneffectivelydecomposehigh-leveltasksintomid-levelplanswithoutanyfurthertraining;ExtrospectiveReasoning

Challenge:foundationmodelsarenotembodiedorgroundedtothephysicalworld

Solution:constrainthemodeltoproposenaturallanguageactionsthatarebothfeasibleandcontextuallyappropriate;ExtrospectiveReasoning

InnerMonologue[1]:injectinginformationfromvarioussourcesoffeedbackinto