OpenAI：2025基于GPT-5的科学研究加速早期实验研究报告（英文版）（89页）.pdf

1、Early science acceleration experiments with GPT-5Sbastien Bubeck1,Christian Coester2,Ronen Eldan1,Timothy Gowers3,Yin Tat Lee1,Alexandru Lupsasca1,4,Mehtaab Sawhney5,Robert Scherrer4,Mark Sellke1,6,Brian K.Spears7,Derya Unutmaz8,Kevin Weil1,Steven Yin1,Nikita Zhivotovskiy91OpenAI2University of Oxfor

2、d3Collge de France and University of Cambridge4Vanderbilt University5Columbia University6Harvard University7Lawrence Livermore National Laboratory8The Jackson Laboratory9University of California,BerkeleyNovember 20,2025AbstractAI models like GPT-5 are an increasingly valuable tool for scientists,but

3、 many remain unawareof the capabilities of frontier AI.We present a collection of short case studies in which GPT-5produced new,concrete steps in ongoing research across mathematics,physics,astronomy,computer science,biology,and materials science.In these examples,the authors highlight howAI acceler

4、ated their work,and where it fell short;where expert time was saved,and wherehuman input was still key.We document the interactions of the human authors with GPT-5,as guiding examples of fruitful collaboration with AI.Of note,this paper includes four newresults in mathematics(carefully verified by t

5、he human authors),underscoring how GPT-5can help human mathematicians settle previously unsolved problems.These contributions aremodest in scope but profound in implication,given the rate at which frontier AI is progressing.ContentsIntroduction2IIndependent rediscovery of known results at the scient

6、ifi c frontier3I.1Improved step-size condition in a recent convex optimization result SbastienBubeck.3I.2Discovering new black hole symmetries with GPT-5 Alex Lupsasca.7I.3Mechanistic analysis and outcome prediction for in vitro immune system exper-iments using GPT-5 Pro Derya Unutmaz,M.D.11II Deep