通过反向传播语言模型反馈来优化生成式 AI.pdfVIP

OptimizinggenerativeAIbybackpropagating

languagemodelfeedback

/10.1038/s41586-025-08661-41,4 ✉1,42,42,42,4

MertYuksekgonul,FedericoBianchi,JosephBoen,ShengLiu,PanLu,

2,4,CarlosGuestrin1,3JamesZou1,2,3 ✉

ZhiHuang

Received:12June2024

Accepted:16January2025

Recentbreakthroughsinartifcialintelligence(AI)

areincreasinglydrivenby

Publishedonline:19March2025

systemsorchestratingmultiplelargelanguagemodels(LLMs)

andotherspecialized

Checkforupdatestools,suchassearchenginesandsimulators.Sofar,thesesystemsareprimarily

handcraftedbydomainexpertsandtweakedthroughheuristicsratherthanbeing

automaticallyoptimized,presentingasubstantialchallengetoacceleratingprogress.

Thedevelopmentofartifcialneuralnetworksfacedasimilarchallengeuntil

backpropagationandautomaticdiferentiationtransformedthefeldbymaking

optimizationturnkey.Analogously,hereweintroduceTextGrad,aversatile

frameworkthatperformsoptimizationbybackpropagatingLLM-generatedfeedback

toimproveAIsystems.Byleveragingnaturallanguagefeedbacktocritiqueand