i.xzea—caeüaxa? 1

i.i.Am@s:—6Ate tgXtB 1

1.2.iii,*,DeepSeek 3

DeepSeek@F?ègê;G—6?;ttg 5

DeepSeek@t+?L?ns?j 5

6

DeepSeek@X@?€(IR 10

11

14

14

14

DeepSeek@gF6—R]Z/g 15

15

15

15

16

20

21

21

21

21

22

2025&1@,DeepSeek-R1?g?Y?, @W1&WI@nGzx?4@?,.,T?X E??@Transformer?1@@???1,

20228°11@

2022°12@

20238°03@

20238°04@

20238°07@

20238°09@

OpenAl??Y??

StabilityAl??,f?1^

OpenAl??N

MidJourneyf?t!?j

MetaQ?f

Anthropic&C

ChatGPT(GPT-3.5),

fi1fit?gü,Stable

GPT-4,@?Bü.@

V5??kzlz(??g@

LLaMA2(oV?￥?/?]

Claude2(?tEJ21

?IáAIGC?g%

Diffusion2.0

????tlfln???kfi

Glf?E)

A?gü)

TR?gE?)

20228o

20248°06@

KimiChatH@?g&?(Q200

20258o

2

) f2e1811,?ta.Xt\g@tZ

?a

ü

%%DEIITzt#gJ,I.UNIXES

OsDeepSeek.DeepSeek-R1:lncentivizingReasoningCapabilityinLLMsviaReinforcementLearning[R].2025.

iü'E.a6T,}@fitIg@%,§üDeepSeekV3 &fA*?g*6710?z,?fi!@Ttoken}?t?@BéQ,BJü370?Z&,

@DeepSeek.DeepSeekMoE:TowardsUltimateExpertSpecializationinMixture-of-EXpertsLanguageModels[R].2024.DeepSeek.DeepSeek-V2:AStrong,Economical,andEfficientMiXture-of-ExpertsLanguageModel[R].2024.

11

s2-10iiifssvlle*As°

@Gou,J.,Yu,B.,Maybank,S.J.etat.KnowledgeDistillation.ASurvey[J].InternationalJournalofComputerVisio,2021.@DeepSeek.DeepSeek-V3TechnicalReport[R].2024.

12

°BROKEtgit?t8

BROKE@TiTét§@}g% (Background)@@(Role)PART(Objectives)X@?*@(KeyResult) 3@@f(Evolve)

B

Background%

?a

R Role@@

0 Objectives@tT

17

°COASTtgifkt8

COAST@iTé§@Tg1TQ(Context) @gfi(Objective) \1fit(Action) W@(Scenario) EN (Task)6

°.f>,@T8mDp0Jt0R3-3@ :

0 Objective@@

A

T Task{I@

TaskEN

T

A

G Goal@f

18

Role@@

Input\zJA

StepsT3fi

E

Expectiong6@

R3-6TRACE}@i j2t§

T TaskEN