人工智能论文英文版-Token Signature： Predicting Chain-of-Thought Gains with Token Decoding Feature in Large Language Models.pdfVIP

TokenSignature:PredictingChain-of-ThoughtGainswithTokenDecoding

FeatureinLargeLanguageModels

PeijieLiu1FengliXu1YongLi1

AbstractincorporatingexamplesofCoTreasoninginafew-shot

prompt(Weietal.,2022),CoTcanbeeffectivelytriggered,

Chain-of-Thought(CoT)techniquehasproven

andtheabilityofLLMstosolvevariouscomplexproblems

5effectiveinimprovingtheperformanceoflargeisimprovedbydecomposingtheproblemstepbystep(Wang

2languagemodels(LLMs)oncomplexreasoningetal.,2022a),whilealsoprovidingdetailedandinterpretable

0tasks.However,theperformancegainsare

2inconsistentacrossdifferenttasks,andtheexplanations(Lanhametal.,2023).InspiredbyCoTreason-

ing,OpenAI(2024a)hasintroducedtheconceptoftest-time

nunderlyingmechanismremainsalong-standing

scaling(Xuetal.,2025),whichsuggeststhatthereasoning

uresearchquestion.Inthiswork,wemakea

JcapabilitiesofLLMscanbeenhancedwithmoretimespent

preliminaryobservationthatthemonotonicityof

6tokenprobabilitydistributionsmaybecorrelatedthinking(test-timecompute).Formanyproblemssuchas

]withthegainsachievedt