q10/random_k4/random_2.md

In-Context Reinforcement Learning for Tool Use in Large Language Models YaoqiYe1 YiranZhao2 KeyuDuan1 ZeyuZheng3 KenjiKawaguchi1 CihangXie4 MichaelQizheShieh1

			Abstract				1.Introduction
							Recent advances		in large	language		models	(LLMs)	(Guo
etal.,2025;Yangetal.,2025;Seedetal.,2025;Teametal.,
While	large	language		models	(LLMs)	exhibit
-----	-----	--------	---	------	------	-------	---	---	---	---	---	---	---	---
2025)haveshowntheireffectivenessinaddressingawide
strongreasoningabilities,theirperformanceon
6202 raM 9 ]IA.sc[ 1v86080.3062:viXra
complextasksisoftenconstrainedbythelimita- range of complex tasks (Wang et al., 2024a; Hsiao et al.,
							2025;Shietal.,2025;Quetal.,2025).					Nevertheless,akey
------------------------------	---	-----------	---	--------------	-----------	----------	---------------------------------	--------	------	------------	----	-----------------	------------	---
tionsoftheirinternalknowledge.					Acompelling
							limitation	remains:	these	models	rely	on	a fixed body	of
approach		to overcome		this challenge		is to aug-
							knowledge	acquired	during	pretraining,		which	inherently
mentthesemodelswithexternaltools—suchas
							restricts their	ability	to	adapt	to new	or time-sensitive		in-
------	------------	---	---	------------	---	--------	---------------	-------	---	-----	------	-----------------	---	---
Python	interpreters		for	mathematical		computa-
formation(Gaoetal.,2023;Zhuetal.,2025;Wangetal.,
tionsorsearchenginesforretrievingfactualin-
							2024b; Cheng	et	al., 2024;	Matarazzo		&	Torlone,	2025).
----------	---	--------	---	--------	------	------	------------	---	----------	---------	---	---	--------	------
formation.		However,		enabling	models	to use
Tomitigatethisissueandenhancemodelflexibility,recent
thesetoolseffectivelyremainsasignificantchal-
researchhasfocusedonenablingLLMstointeractwithex-
lenge.	Existing	methods		typically	rely	on cold-
------	---------	-------	-----	---------	----------	--------	---------------------------	---	---	-------------------------	---	---	---	---
							ternaltoolsduringinference.			Thisincludesgeneratingand
start	pipelines	that	begin	with	supervised	fine-
executingPythoncodeformathematicalreasoning,lever-
tuning(SFT),followedbyreinforcementlearning
agingwebsearchenginestoaccessup-to-dateanddomain-
(RL).Theseapproachesoftenrequiresubstantial
specificcontent,andinvokingdedicatedhelpermodelsfor
amounts	of	labeled	data	for SFT,	which	is ex-
-------	---	-------	----	--------	-----	------	---	---	---	---	---	---	---	---
pensivetoannotateorsynthesize. Inthiswork, specializedsubtasks(Guoetal.,2024;Teametal.,2025;Li
etal.,2025b;Jinetal.,2025a;Fengetal.,2025).
weproposeIn-ContextReinforcementLearning
(ICRL),	an	RL-only	framework		that	eliminates
-------	---	-------	---------	---	----	----------	------------	--------	---	---------	---	----	------	------
							The dominant	training		paradigms	for	LLMs	either	lever-
theneedforSFTbyleveragingfew-shotprompt- agereinforcementlearning(RL)withverifiablerewardsig-
ing during the rollout stage of RL. Specifically, nals(Guoetal.,2025;Jinetal.,2025a;Zhaoetal.,2025),
ICRLintroducesin-contextexampleswithinthe
							or adopt	a cold-start	strategy		that begins	with	supervised
---	---	---	---	---	---	---	--------	------------	--------	---	-----------	----	----------	---
rollout prompts to teach the model how to in- fine-tuning(SFT)followedbyanRLphase(Meietal.,2025;
voke	external	tools.	Furthermore,		as	training
----	--------	------	------------	---	---	--------	------------------	---	-----------------------------	---	---	---	---	---
							Nguyenetal.,2025).		DirectlyapplyingRLfromscratch
progresses, the number of in-context examples often yields poor performance, as the model lacks initial
isgraduallyreduced,eventuallyreachingazero- tool-useabilitiesandstruggleswithineffectiveexploration.
shotsettingwherethemodellearnstocalltools
WhileincorporatinganSFTstagecanguidethemodelto-
independently.Weconductextensiveexperiments wardamorefavorableinitialization,ittypicallyrequiresa
acrossarangeofreasoningandtool-usebench- largeamountofhigh-qualitylabeleddata,whichisexpen-
ICRL
marks. Results show that achieves state- sivetoannotateorsynthesize.
of-the-art		performance,		demonstrating		its effec-
----------	---	------------	---	-------------	---	----------	-------------	---	------------	---	----------	---	-------------	---
							In this work,		we introduce		In-Context		Reinforcement
tivenessasascalable,data-efficientalternativeto
					1		Learning	(ICRL),	a lightweight		and	supervision-efficient
---	---	---	---	---	---	---	--------	-------	-------------	---	---	---------------------	---	---
traditionalSFT-basedpipelines.
frameworkfortrainingLLMstoperformtool-augmented
							reasoning.	UnlikepriorapproachesthatrelyonSFT,ICRL
---	---	---	---	---	---	---	----------	---------------------------------------	---	---	---	---	---	---
teachestoolusedirectlythroughRLrolloutsthatareaug-
*Equal contribution 1National University of Singapore mentedwithin-contextdemonstrations. Specifically,during
2Salesforce	AI	Research	3University		of California	Berkeley
-----------	---	--------	-----------	---	-------------	--------	---	---	---	---	---	---	---	---
RLtraining,weconstructeachrolloutpromptbyprepending
4University	of California,		Santa	Cruz.	Correspondence	to: Yi-
-----------	--------------	---	-----	-----	--------------	-------	---	---	---	---	---	---	---	---
asmallnumberoffew-shotexamplesthatillustratehowto
ran Zhao	zhaoyiran0924@gmail.com,				Michael	Qizhe Shieh
--------	--------------------------	---	---	---	-------	-----------	---	---	---	---	---	---	---	---
michaelshieh@comp.nus.edu.sg. reasonstep-by-step,invoketoolsinastructuredformat,and
							generatefinalanswers.		Thesedemonstrationsserveassoft
---	---	---	---	---	---	---	---------------------	---	------------------------------	---	---	---	---	---
Preprint.March10,2026. supervisionduringexploration,guidingthemodeltoward
1Code				https://github.com/
-----	---	---	---	-------------------	---	---	---	---	---	---	---	---	---	---
is publicly available at successfulbehaviorwithoutrequiringlabeledtrajectories.
applese233/ICRL
1

In-ContextReinforcementLearningforToolUseinLargeLanguageModels Figure1.ICRLtrainingworkflow.Themodelistrainedthroughamulti-stagecurriculumthatgraduallyreducesthenumberofin-context examplesintherollouttemplate.Ateachstage,theLLMgeneratestool-augmentedrollouts,receivesrewards,andupdatesitspolicyvia reinforcementlearning,enablingatransitionfromimitationtoautonomoustooluse. Furthermore,astrainingprogresses,wegraduallyreducethe TheseresultshighlightICRL’sabilitytogeneralizetodi- numberofdemonstrationsincludedintheserolloutprompts, versetool-augmentedreasoningdomainsanditspotentialas transitioningthemodelfromfew-shottozero-shotsettings. aunified,scalableframeworkfortrainingtool-usingmodels Thisprogressivereductionformsacurriculumthatencour- withoutcostlysupervision. agesthemodeltointernalizetool-usestrategiesandproduce structuredoutputsautonomously,withoutrelyingonprompt- 2.In-ContextReinforcementLearning(ICRL)

basedscaffolding.	WeoptimizethemodelusingRLwitha
rewardthatbalancestaskaccuracyandformatcorrectness. In this section, we formally introduce tool use in LLMs
Toensurestability,weadoptGRPO(Shaoetal.,2024)with anddescribehowRLcanbeappliedtotrainsuchbehavior.
lossmaskingtoignorenon-trainabletooloutputs. Byem- WealsopresenttheoverallworkflowofICRL,detailingits
beddingandgraduallyremovingdemonstrationsfromthe trainingtemplates,learningprocess,andrewarddesign.
RLrollouts,ICRLmergestheefficiencyofpromptingwith
theadaptabilityofRL,offeringascalable,supervision-light
2.1.ToolUseinLLMs
alternativetotraditionalSFT+RLpipelines.
				When	LLMs encounter	queries	that exceed	the	scope of
---	---	---	---	----	--------------	-------	-----------	---	--------
Weconductcomprehensiveexperimentsacrossarangeof
theirinternalknowledge,theymustleverageexternaltools
QA and reasoning benchmarks to evaluate the effective- to obtain updated information or perform more complex
nessofICRL.Withoutrelyingonsupervisedfine-tuningor
				reasoning.	Forexample,searchenginescanprovideaccess
---	---	---	---	----------	----------------------------------------	---	---	---	---
ground-truthtooltraces,ICRLachievesstate-of-the-artper- torecentknowledge,whilePythoninterpreterscanbeused
formanceonchallengingQAdatasets,outperformingstrong toexecutestructuredreasoningprocedures.
baselines such	as ZeroSearch	(Sun et al.,	2025), Search-
--------------	-------------	------------	--------------	---	---	---	---	---	---
Formally,givenaqueryqandanexternaltoolT,themodel
R1(Jinetal.,2025a),andParallelSearch(Zhaoetal.,2025)
				generates	a response	y =	(y ,y ,...,y	), where	each
----------------------------------------------	---	---	---	---------	--------------	--------	------------	---------	--------
byupto8.9onQwen2.5-3B(Yangetal.,2024b)and7.3on							1 2		y
				token	is conditioned	not only	on the	query and	previous
Qwen2.5-7Binaverageexactmatchaccuracy.Thegainsare
tokens,butalsoonahistoryofpriorinteractionswiththe
especiallypronouncedonmulti-hopreasoningtasks,where
ICRLachievesdouble-digitimprovementsondatasetslike tool. Thisdefinesaconditionaldistributionoftheform:
TriviaQA(Joshietal.,2017),2Wiki(Hoetal.,2020),and							y
-------------------------------------------------	---	---	---	---	---	---	---	---	---
Musique(Trivedietal.,2022). Furthermore,incontrastto (cid:89)
					π θ (y	q,T)=		π θ (y t	y
---	---	---	---	---	-------------	---	-----------	---------	-----
methodssuchasO2-Searcher(Meietal.,2025)thatrequire
t=1
cold-start SFT	to learn complex	tool-use	behavior, ICRL
--------------	----------------	--------	--------------	---	---	---	---	---	---
learnssuchcapabilitiesdirectlythroughin-contextexamples Here,π θ isthemodelparameterizedbyθ,andH t denotes
				the sequence	of previous	actions	taken	by the	model and
---	---	---	---	------------	-----------	-------	-----	------	---------
duringRLrollouts—demonstratingsuperiordataefficiency.
thecorrespondingobservationsreturnedbythetoolupto
BeyondwebQA,wealsoevaluateICRLonmathreasoning
				step t.	Specifically,	the interaction		between	the model
-----------------------------------	-----------	-------------	----------	-------	------------------	---------------	----------	-----------	---------
tasksinvolvingcodeexecutionasatool.		OntheAIME2024
				and the	tool is structured	as	a sequence	of actions.	At
and AIME2025	benchmarks,	ICRL matches	or exceeds
the performance of ReTool (Feng et al., 2025), a strong each time step, the model may choose to (i) perform in-
ternalreasoning,(ii)issueaquerytotheexternaltool,or
SFT+RLbaseline,despiteusingnosupervisedpretraining.
				(iii) return	a final	answer.	These actions	are	embedded
---	---	---	---	------------	-------	-------	-------------	---	--------
2

In-ContextReinforcementLearningforToolUseinLargeLanguageModels

inthegeneratedtextinastructuredformat,suchasXML							where
tags,	which	distinguish	reasoning	steps	from	tool invoca-
							R(τi)−mean({R(τi)	τi∼π			(τ),i=1,2,...,N})
θold
tions and answers. For example, a reasoning step might Ai= , (4)
								std({R(τi)	τi∼π		θold (τ),i=1,2,...,N})
---	-------	----------------------	---	---	--------	-----	---	---------------	---	----------------------	---	---	---
be	denoted	as ...,			a search	query
as ..., a retrieved information andr (θ)=π (τ	q,τ )/π (τ	q,τ ).
							i,t		θ i,t	i,<t θold	i,t i,<t
---	---	---	---	---	---	---	---	---	-----	---------	--------	---	---
as...,andafinal
answeras....							2.3.ICRL
-----------------------------	---	---	---	---	---	---	--------	---	---	---	---	---	---
Furthermore,thetoolfunctionsasaresponsemechanism. Training Process. Rather than training models from
Forexample,asearchenginecanbemodeledasaretrieval
scratchusingreinforcementlearning,whichoftensuffers
		V∗	V∗,	V∗
---	---	---	---	---	---	---	---	---	---	---	---	---	---
function T : → where is the space of textual fromsparserewardsandinefficientexploration,orrelying
sequences. Givenasearchqueryq′,thetoolreturnsanob- exclusivelyonfew-shotprompting,whichincurssubstantial
servationo=T(q′),suchasthetop-kdocumentsretrieved
inferenceoverhead,weintroduceICRL,aframeworkthat
fromacorpus. Thisobservationisappendedtothemodel’s integrates the strengths of both approaches. ICRL lever-
contextandusedinsubsequentgenerationsteps.
agesthesampleefficiencyandinductivebiasoffew-shot
promptingwhilebenefitingfromtheexplorationcapabilities
2.2.RLwithToolUse							ofreinforcementlearning.
-----------------	---	---	---	---	---	---	------------------------	---	---	---	---	---	---
RL Objective Function. After formulating the tool- Atthebeginningoftraining,weincorporateasmallnumber
augmentedreasoninginLLMsasaMarkovDecisionPro- of tool-use demonstrations into the model’s rollout tem-
cess(MDP),wecandefinethecorrespondingreinforcement plate. These examples guide the model toward effective
learning(RL)objectiveasfollows: tool-augmentedreasoningviain-contextlearning,akinto
							few-shotprompting.		Theresultingpolicyisdenotedas:
---	---	---------------	---	------	---	---	------------------	---	------------------------------	---	---	---	---
	max	E	[r	(q,y)]
		q∼D,y∼πθ(·	q,T)		ϕ
πθ
						(2)					y
---	---	---	-------------	---	----------	---	----	---	------	--------	-------	---	---
		−βD	[π (y	q,T)∥π		(y	q,T)],					(cid:89)
		KL	θ	ref			π (y		P	,q,T)=	π (y	P	,y ,q,H
							θ	N		θ t	N <t	t
t=1
whereπ		isthepolicyLLM,π		isthereferenceLLM,r
------	---	----------------	---	-------------------	---	---	---	---	---	---	---	---	---
	θ			ref		ϕ
istherewardfunctionandD isKL-divergencemeasure. whereP N representsthefew-shotpromptconsistingofN
KL
							demonstrationexamples.			Table1showsaconcreteexample
---	---	---	---	---	---	---	----------------------	---	---	---------------------------	---	---	---
ofrollouttemplate.
LossMasking.		UnliketraditionalRL,whichoptimizes
------------	---	----------------------------------	---	---	---	---	---	---	---	---	---	---	---
solely over model-generated tokens, tool-augmented rea- Aftertrainingforseveralsteps,themodelbeginstoacquire
soning introduces retrieved content into the rollout se- tool-usecapabilitieswiththeguidanceoftheinitialfew-shot
quence—tokens that are not produced by the model and promptP . Oncesufficientlearningprogressisobserved,
N
therefore do not reflect its internal reasoning or decision- wepausetrainingandreducethenumberofdemonstration
makingprocess.		Toaddressthis,weadoptalossmasking
--------------	---	---------------------------------	---	---	---	---	--------------------	---	---	-----------------------------	---	---	---
							examplesintheprompt.			Theupdatedpolicyconditionedon
strategytailoredforRLwithtooluse, whichexcludesre- areducedpromptP isdefinedas:
N−1
trieved	content	from	the optimization.		Specifically,	only
------------------------------------------------	-------	----	-----------------	---	-------------	----	---	---	---	---	---	---	---
tokensgeneratedbythelanguagemodelcontributetothe											y
(cid:89)
							π (y	P		,q,T)=	π (y	P	,y
---	---	---	---	---	---	---	-------	---	------	-------	---	----	------
policygradient,whileretrievedspansaremaskedoutand θ N−1 θ t N−1 <t t
excludedfromthelosscomputation.				Thistargetedoptimiza-						t=1
-------------------------------	---	---	---	---------------------	---	---	---	---	---	---	---	---	---
tionensuresthatlearningremainsfocusedonthemodel’s whereP denotesapromptwithN −1demonstration
N−1
ownbehavior—suchastoolusage,intermediatereasoning,
							examples.	Thisprocessisrepeatediteratively,progressively
---	---	---	---	---	---	---	---------	----------------------------------------------	---	---	---	---	---
andfinalanswers—withoutbeingaffectedbyfixed,untrain-
reducingthenumberofdemonstrations,untilnoexamples
ablecontentfromexternalsources.							remainintheprompt.
-------------------------------	---	---	---	---	---	---	------------------	---	---	---	---	---	---
GRPOwithToolUse. WeadoptGRPO(Shaoetal.,2024) RewardDesign. Wedesignacompositerewardfunction
totrainπ		ontheRLdatasetD		={q ,q	,··· ,q	}. Specif-
--------	---	---------------	---	------	-------	----------	-------------------------------------------------	---	---	---	---	---	---
	θ			1	2	n	thatcombinestheansweraccuracyandformatcorrectness
ically,forq ∈D,weusetheoldpolicyfrompreviousstep toprovidearicherlearningsignal:
π	tosampleagroupofN		individualresponsesτ			. Then,
---------------------	-----------------	---	--------------------	---	--------	--------	----------------	------	--------------	-------------	-------	------------	-----
θold						i
theRLlossisdefinedas:							r (q,y)=α·reward			+(1−α)·reward			, (7)
							ϕ			acc		format
							τi		where α	is the	hyperparameter	to	balance
				1	N
	(θ)=E				(cid:88)	(cid:88)
L GRPO Specifically, the accuracy-based reward is computed us-
		τi∼πθold	(q),q∼DRL(cid:80)N			τ		(3)
---	---	--------	------------------	---	----	---	---	---	---	---	---	---	---
i=1 i i=1 t=1 ing exact match (EM) between the model’s predicted an-
				,ϵ)−β·D			swer and	the	ground	truth. The reward	is	assigned	as
---	---	------	---------	-------	-----	-----------	--------	---	------	-----------------	---	--------	---
		CLIP(r	i,t (θ),A	i	KL [π	θ ∥π ref ],
3

In-ContextReinforcementLearningforToolUseinLargeLanguageModels Table1.Few-shotrollouttemplateinICRL. Few-ShotPromptTemplate Solvethefollowingproblemstepbystep. Youmustconductreasoninginside...everytimeyougetnew information.Afterreasoning,ifyoufindyoulacksomeknowledge,youcancallasearchengineby

query anditwillreturnresultsbetween....Youcansearchasmanytimesasyouwant.Finally, providetheanswerinside.... Herearesomeexamples: ExampleProblem:q demo ExampleSolution:...... ......a

(repeatedforN		examples
Nowsolvethefollowingproblem:
ActualProblem:question
Table2.Formatviolationpenaltiesforcomputingreward					.
-------------------------------------------------	---	---	---	---	---	---	---	---	---	---	---	---
format
Algorithm1ICRL
Violation			Rationale
---------	---	---	---------	---	---	------	-------	--------	-----------	-----	---	------
						Input:	Initial	policy π	, reference	model	π	, tool
								θ			ref
Notag Mustprovidestructuredanswer T, initial few-shot prompt P , dataset partitions
N
Unbalancedtags ProperXMLstructurerequired {D(N),D(N−1),...,D(0)}, reward function r (·),
ϕ
Notag			Shoulddemonstratereasoning
---------------------	---	---	--------------------------	---	---	----------------	---	---	---	---	---	---
Unbalancedtags						numberofRLstepsT
ProperXMLstructurerequired
Nousage Shouldutilizeavailabletool Output: Trainedmodelπ θ
Emptyanswercontent			Answermustbesubstantive			1: fork=N		to0do
------------------	---	---	-----------------------	---	---	---------	-------	------------	---	------	----	---
						2:	// Step	1: Construct		prompt	with
k demonstrations
P ←selectkexamplesfromP
reward	= 1 if the	prediction	exactly	matches	the cor-	3:	k			N
-------------------------	----------	----------	-------	-------	--------	---	------------------------------------------	---	---	---	---	---
	acc						D(k)←RLtrainingsubsetforcurrentpromptlevel
rectanswer,and0otherwise.						4:
						5:	fort=1toT	do
The reward component evaluates the model’s adher- 6: forq∈D(k)do
format
encetotheexpectedstructuredoutputformat,specifically						7:	π	←π
----------------------------------------------------	---	---	---	---	---	---	---	---	---	---	---	---
θold θ
						8:	Sample	N trajectories		{τ	,...,τ	} ∼
-----------------------	---	---	--------------	---	---	---	------	--------------	---	---	------	---
thecorrectuseofXMLtags.			Itisdefinedas:							1	N
π (q,P ,T)
θold k
			(cid:88)				foreachtrajectoryτ		do
---	------	-----	-----------	---	---	---	------------------	---	---	---	---	---
	reward	=1.0−	penalty(v),		(8)	9:			i
format
						10:		Computereward:r	ϕ	(q,τ i )
---	---	---	---	---	---	---	---	---------------	---	--------	---	---
v∈V
						11:		ComputenormalizedadvantageA			i
---	---	---	---	---	---	---	---	---------------------------	---	---	---	---
whereV denotesthesetofformatviolationsidentifiedinthe 12: Computeimportanceweightsr (θ)
i,t
						13:	endfor
----------------	---	-----------------------------------	---	---	---	---	--------------	---	---	----	---	---
model’sresponse.		Thepenaltyfunctionpenalty(v)assigns
						14:	Updatepolicy:π		←π	−∇ L
apredefinedcosttoeachviolation,asspecifiedinTable2. θ θ θ GRPO
endfor
15:
Withtheproposedrewarddesign,weoptimizethepolicy						16:	endfor
-----------------------------------------------	---	---	---	-----------	---	----------	------	---	---	---	---	---
usingtheRLobjectivedefinedinEquation3.				Thecomplete		17: endfor
trainingprocedureforICRLisoutlinedinAlgorithm1.
						convergenceduringRLtraining.			Forimprovedtrainingeffi-
---	---	---	---	---	---	----------------------------	---	---	------------------------	---	---	---
3.Experiment ciency,allmodelsareloadedusingbfloat16precision.
3.1.Setup
WeevaluatetheeffectivenessofICRLbycom-
Baselines.
Backbone Models. We apply ICRL to the Qwen2.5 paringitagainstseveralstate-of-the-artmethodsfortraining
model family (Yang et al., 2024a), focusing primarily on tool-augmentedLLMs. Thesebaselinesfallintothreemain
Qwen2.5-3B-InstructandQwen2.5-7B-Instruct,andfurther categories. Directpromptingmethodsincludemodelsthat
evaluatingonQwen2.5-14B-Instruct. Wealsoextendour performinferenceusingdirectinputsorpromptingstrate-
experimentstotheQwen3series(Yangetal.,2025),particu- giessuchasChain-of-Thought(CoT)reasoning(Weietal.,
larlyQwen3-8B,whichincorporatesRLenhancements. All 2022). Retrieval-basedmethodsleverageexternalinfor-
theseinstruction-tunedmodelsarewidelyadoptedforques- mationthroughtechniqueslikeRetrieval-AugmentedGen-
tionansweringandreasoningtasks. Wechoosetheinstruct eration(RAG),includingstandardRAG(Lewisetal.,2020),
variants over base models due to their strong instruction- InterleavingRetrievalChain-of-Thought(IRCoT)(Trivedi
followingcapabilities,whichenablefasterandmorestable etal.,2023),andSearch-o1(Lietal.,2025a). Fine-tuning-
4

In-ContextReinforcementLearningforToolUseinLargeLanguageModels Table3.MainResultsofICRL:ExactMatch(EM)Accuracy(%)onvariousdifficultQAdatasets.Thebestperformanceissetbold.The secondbestperformanceisunderlined. DifficultQuestionAnswering

		Model	Method								Average
					TriviaQA	HotpotQA	2Wiki	Musique	Bamboogle
			Direct		28.8	14.9	24.4	2.0		2.4	14.50
			CoT		3.2	2.1	2.1	0.2		0.0	1.52
			IRCoT		31.2	16.4	17.1	6.7		24.0	19.08
			Search-o1		47.2	22.1	21.8	5.4		32.0	25.70
			RAG		54.4	25.5	22.6	4.7		8.0	23.04
			SFT		29.2	18.6	24.8	4.4		11.2	17.64
Qwen2.5-3B
			R1-instruct		44.9	20.8	27.5	6.0		19.2	23.68
---	----------	---	--------------	---	----	----	----	----	---	----	----------	---
			RejectSampling		48.8	24.0	23.3	5.9		21.0	24.60
			Search-R1		54.5	32.4	31.9	10.3		26.4	31.10
			ZeroSearch		57.4	27.4	30.0	9.8		11.1	27.14
			ICRL		72.6	35.4	39.2	20.0		33.6	40.16+8.94
			Direct		40.8	18.3	25.0	3.1		12.0	19.84
			CoT		18.5	9.2	11.1	2.2		23.2	12.84
			IRCoT		47.8	13.3	14.9	7.2		22.4	21.12
			Search-o1		44.3	18.7	17.6	5.8		29.6	23.20
			RAG		58.5	29.9	23.5	5.8		20.8	27.70
			SFT		35.4	21.7	25.9	6.6		11.2	20.16
	Qwen2.5-7B		R1-base		53.9	24.2	27.3	8.3		29.6	28.66
			R1-instruct		53.7	23.7	29.2	7.2		29.3	28.62
			RejectSampling		59.2	33.1	29.6	12.3		35.5	33.94
			Search-R1		61.0	37.0	41.4	14.6		36.8	38.16
			ZeroSearch		65.2	34.6	35.2	18.4		27.8	36.24
			ParallelSearch		62.8	42.9	42.4	19.7		41.1	41.78
			ICRL		75.4	42.6	53.6	26.0		48.0	49.12+7.34
based methods involve approaches such as SFT (Chung EvaluationBenchmarks. WeevaluateICRLandvarious
et al., 2024), RL without search (R1) (Guo et al., 2025), baselinesonseveralwidely-usedQAbenchmarks,including
andRejectionSampling(Ahnetal.,2024). Wealsoinclude TriviaQA(Joshietal.,2017),HotpotQA(Yangetal.,2018),
recentRLmethodsthatintegratesearchcapabilities,suchas 2Wiki (Ho et al., 2020), Musique (Trivedi et al., 2022),
Search-R1(Jinetal.,2025a),ZeroSearch(Sunetal.,2025), andBamboogle(Pressetal.,2023). Sinceourmodelsare
O2-Searcher(Meietal.,2025),andParallelSearch(Zhao
trainedontheNaturalQuestions(NQ)dataset,weexclude
et al., 2025). These baselines provide a comprehensive NQfromtheevaluationtoavoiddataleakage. Thesebench-
comparisontovalidatethegeneralityandadvantagesofour markscoverdiversedomainsandreasoningtypes,providing
proposedICRLframework. acomprehensiveassessmentofmodelperformance.Further-
more,toensureevaluationefficiency,werandomlysample
Training	Datasets.		We use	the Natural	Questions (NQ)
--------	---------	---	------	-----------	--------------	---	--------------------------------	---	---	---	-----------------	---
							upto500questionsfromeachdataset.				Theselectedbench-
dataset(Kwiatkowskietal.,2019)astheprimarytraining
marksincludebothin-domaingeneralQAtasks(e.g.,Triv-
corpus.	The	dataset	is loaded	via FlashRAG	(Jin et	al.,
-------	---	-------	---------	------------	-------	----	---	---	---	---	---	---
iaQA,HotpotQA)andout-of-domainmulti-hopQAtasks
2025b),whichprovidespreprocessedquestion-answerpairs
(e.g.,2Wiki,Musique,andBamboogle),allowingustothor-
withgold-standardanswers.				NQcontainsrealuserqueries
-------------------------	---	---	---	-------------------------	---	---	---	---	---	---	---	---
oughlytestthegeneralizationandreasoningcapabilitiesof
fromGoogleSearch,eachpairedwithWikipediapassages
differentmethods.
that	includethe	correct	answer.	To support	ourproposed
----	----------	-------	-------	----------	-----------	---	---	---	---	---	---	---
trainingmethod,werandomlysampledthreequestionsfrom
thewebandusedGPT-5.22togeneratefew-shotexamples Reward. The hyperparameter α in Equation 7 is set to
formattedaccordingtotherollouttemplateshowninTable1. 0.8. Furthermore,fortheformatviolationpenaltiesinTa-
Tosimulatereal-worldtool-usebehavior,weintegratethe ble2,theweightsaresetto0.5,0.2,0.15,0.1,0.1,and0.2
SerperAPI3							respectivelyfromtoptobottom.
----------	---	---	---	---	---	---	----------------------------	---	---	---	---	---
acrossallmodelstoretrieveliveresultsfrom
theGoogleSearchengine.			Forfairness,eachqueryretrieves
----------------------	---	---	------------------------------	---	---	---	---	---	---	---	---	---
thetop3documentsrequiredforsearch-basedreasoning. ImplementationDetails. Weimplementourmethodus-
							ing the Volcano	Engine	Reinforcement		Learning	(VeRL)
---	---	---	---	---	---	---	---------------	------	-------------	---	--------	------
2https://platform.openai.com/docs/models/
gpt-5.2							framework	(Sheng	et al.,	2024).	Qwen2.5-3B-Instruct,
-------	---	---	---	---	---	---	---------	------	-------	------	--------------------	---
3https://serper.dev/ Qwen2.5-7B-Instruct,andQwen2.5-14B-Instructserveas
							the backbone	models,	trained	with	a learning	rate of 1e-
---	---	---	---	---	---	---	------------	-------	-------	----	----------	-----------
5

In-ContextReinforcementLearningforToolUseinLargeLanguageModels Table4.Results:ExactMatch(EM)Accuracy(%)onvariousQAdatasets.O2-Searcherappliescold-startSFTbeforeRL,whileour method(ICRL)appliesRLwithoutSFT.

	Model		Method		SFT		DifficultQuestionAnswering				Average
						TriviaQA	HotpotQA	2Wiki	Musique	Bamboogle
			O2-Searcher		✓
						59.7	38.8	37.4	16.0		34.4 37.26
	Qwen2.5-3B				✗
			ICRL			72.6	35.4	39.2	20.0		33.6 40.16
Table5.Exampleofmulti-turnsearch-augmentedquestionanswering.ThemodelcorrectlyidentifiesGeorgeWashingtonasthepresident
whosetthetwo-termprecedentandretrieveshisinaugurationdatethroughtwosearchqueries.
Question Whendidthepresidentwhosettheprecedentofatwotermlimitenteroffice?
	GoldenAnswer			April30,1789
---	------------	---	---	------------	---	---	---	---	---	---	---	---
ModelResponse Ineedtoknowwhichpresidentsetthisprecedent. Ialsoneedtoknowwhenhe
enteredoffice.Letmestartbyfindingoutwhichpresidentitis.
presidentwhosettwotermlimitprecedent
Doc1(Title:“Twenty-secondAmendmenttotheUnitedStatesConstitution”)
Untiltheamendment’sratification,thepresidenthadnotbeensubjecttotermlimits,butboth
GeorgeWashingtonandThomasJefferson(thefirstandthird...Doc2(Title:“Amendment22
–TermLimitsforthePresidency”)ThedecisionforGeorgeWashington–theFirstPresident
of the United States – to voluntarily step down from office after two terms, established an
unofficial...Doc3(Title:“ThefirstU.S.presidentsethisowntermlimit”)Washingtonwrote...
ItseemsGeorgeWashingtonsettheprecedent. NowIneedtofindoutwhenhe
enteredoffice.
whendidGeorgeWashingtonenterofficeaspresident
Doc1(Title: “PresidencyofGeorgeWashington”)Washingtonwasinau-
guratedasthefirstPresidentoftheUnitedStatesonApril30,1789... Doc2(Title: “George
Washington”)Washingtonwasunanimouslyelected... Doc3(Title:“PresidentWashington’s
FirstTerm(1789-1792)”)GeorgeWashingtonwasinaugurated...
Based on the information, George Washington entered office on April 30, 1789.
April30,1789
6. For each query, we sample 8 rollout trajectories with (EM)scoreof40.16,surpassingthebestcompetingmethod,
atemperatureof1.0tocomputethegroup-relativeadvan- Search-R1 (31.10), by +8.94. The improvements are es-
tage. Themaximumpromptlengthissetto5000tokensto peciallypronouncedonmulti-hopdatasetssuchas2Wiki
accommodatefew-shotdemonstrations,andthemaximum (+7.3),Musique(+9.7),andBamboogle(+7.2),demonstrat-
response length is capped at 2048 tokens, allowing up to ing ICRL’s strength in handling complex reasoning and
6searchturnsperquery.			Toregularizethepolicy,				weap- tool-usescenarios.
---------------------	-------	----	----------------------	---	---------	--------	------------------------	---	---	---	---	---
ply a KL	penalty	with	a coefficient		of 0.001.	Training	is
Furthermore,onQwen2.5-7B,ICRLachievesanaverage
conductedon4NVIDIAA100GPUs(80GBeach),using
							EMscoreof49.12,		outperformingthestrongestbaseline,
---------------	--------	-------------------------------	--------	-------------	---	-----------	---------------	---	----------------------------------	---------	-----------	------------
abatchsizeof64.		WeadoptFullyShardedDataParallel
							ParallelSearch		(41.78),	by +7.34.	It achieves	the best re-
(FSDP)	training	with	gradient	checkpointing		to optimize
sultsonfouroutoffivedatasets,includingTriviaQA(75.4),
memoryusage.		Forretrieval,weuseaBM25retrieverthat
------------	---	------------------------------------	---	---	---	---	---	---	---	---	---	---
2Wiki(53.6),Musique(26.0),andBamboogle(48.0).These
returnsthetop-3documentsforeachsearchquery.
resultsshowthatICRLscaleseffectivelywithmodelsize
							and	generalizes	well	across both	in-domain	and out-of-
---	---	---	---	---	---	---	---	-----------	----	-----------	---------	-----------
3.2.MainResults
							domainQAtasks.		Theconsistentgainsoverbaselinesthat
---	---	---	---	---	---	---	--------------	---	-----------------------------------	---	---	---
relyonsupervisedfine-tuningorrewardmodeling—such
Table3presentsthemainresultscomparingICRLtoother
baselines across five popular QA benchmarks. From the asZeroSearch,Search-R1,andRejectSampling—highlight
results,wecanobservethat: theeffectivenessofourin-contextreinforcementlearning
frameworkinlearningtool-usebehaviorswithoutexplicit
accuracy-basedrewardsorsupervision.
ICRL	achieves	state-of-the-art			performance	across
--------------	--------	----------------	--------	---	-----------	----------------	---	---	---	---	---	---
QA benchmarks.			As shown	in	Table	3, ICRL signifi-
ICRL
cantlyoutperformsallbaselinesonbothQwen2.5-3Band achieves better performance without SFT or la-
Qwen2.5-7BmodelsacrossfivechallengingQAdatasets. beleddata. Table4highlightsakeyadvantageofICRL:
On Qwen2.5-3B, ICRL achieves an average exact match it achieves superior performance without requiring any
6

In-ContextReinforcementLearningforToolUseinLargeLanguageModels (a)EMAccuracy (b)FinishPercent 100

						320		100
321~0
80	75.4
-------------	----	---	---	---	---	---	---	-------------------	---	---	---	---	---	---
)%(ycaruccAME								)%(hsiniFevitalumuC
80
60
53.6
						48		60
---	---	---	---	---	---	---	---	---	---	---	---	---	---	---
42.6
40
320
				26.8	26			40
---	----	---	----	----	---	----	---	---	---	---	---	---	-------	---
	20.8												321~0
20			17.8			14.4
20
9
0
	A	A	Wiki		Que mboogle				0
------	--------	---	----	----	-----------	---	---	---	---	---	---	---	---	---
Trivia	Q Hotpot	Q		MuSi
			2						1	2	3	4	5 6	7
Ba
NumberofTurns
Figure2.ComparisonofQwen-7Btrainedforthreestages(320)vsfourstages(321~0).(a)EMAccuracyacrossfiveQAdatasets.
(b)Cumulativefinishpercentvsnumberofsearchturns,aggregatedacrossalldatasets.
Table6.Results:Qwen2.5-14BmodelsExactMatch(EM)Accuracy(%)onvariousQAdatasets.Thebestperformanceissetbold.
DifficultQuestionAnswering
		Model		Method								Average
---	-----------	-----	---	------	--------	----	--------	-----	-------	---------	----	-------	---	---
					TriviaQA		HotpotQA	2Wiki	Musique	Bamboogle
				Direct		52.0	22.6	28.2	6.0		15.2	24.80
	Qwen2.5-14B			CoT		56.4	24.6	25.8	9.0		40.0	31.16
				ICRL		75.0	43.2	61.8	25.6		53.6	51.84
SFT,incontrasttoO2-Searcher,whichappliesacold-start throughin-contextreinforcementlearning.
SFT phase	before	reinforcement		learning.		Despite	using
----------	------	-------------	----------------	---------	------------	-------	-----	---	---	---	---	---	---	---
no labeled	tool	traces	or task-specific		supervision,		ICRL
4.FurtherAnalysis
achievesahigheraverageEMscoreof40.16comparedto
37.26 from O2-Searcher. It outperforms O2-Searcher on 4.1.Ablationanalysis
fouroutoffivedatasets,includingsubstantialgainsonTrivi-
								Ablation	on curriculum		design	for	rollout reduction.
---	---	---	---	---	---	---	---	--------	-------------	---	------	---	------------------	---
aQA(+12.9)andMusique(+4.0).Theseresultsdemonstrate
Weconductanablationstudycomparingtwocurriculafor
thatICRLcanlearneffectivetool-usestrategiespurelyfrom
reducingthenumberofexamplesusedintherolloutprocess:
in-contextexamplesandreinforcementsignals,offeringa
athree-stageschedule(320)andafour-stageschedule
scalableanddata-efficientalternativetomethodsthatrely
								(321~0).	AsshowninFigure2(a),thethree-stagevariant
---	---	---	---	---	---	---	---	----------	-----------------------------------------	---	---	---	---	---
oncostlyannotationandpretraining.
achievessubstantiallyhigherEMaccuracyacrossallfive
								QA datasets.	For	instance,	on	TriviaQA	and 2Wiki,	the
---	---	---	---	---	---	---	---	------------	---	---------	---	--------	----------	---
3.3.ConcreteExamples
three-stagemodelreaches75.4and53.6,comparedto20.8
Table5presentsacompletereasoningexamplefromICRL- and26.8withthefour-stageversion.
Qwen2.5-7B	on	a question		from the	Bamboogle	dataset.
----------	---	----------	---	--------	---------	--------	---	---	---	---	---	---	---	---
Figure2(b)showsthatthefour-stagecurriculumleadsto
Themodelistaskedwithansweringacompositionalquery:
fasterdecisions,withover80%ofqueriesfinishingwithin
identifying	the	president	who	established		the two-term
-----------	---	---------	---	-----------	---	------------	---	---------------	----------------------------------	---	---	---	---	---
								twosearchturns.	However,thiscomesatthecostofanswer
precedentanddeterminingwhenheenteredoffice. Itfirst quality. These results suggest that aggressively reducing
issuesasearchquerytoidentifytherelevantfigure,correctly
								rollout length	too	early	(via the	intermediate	stage	with
----------	----	------	----------	---	-------	----------	---	--------------	---	-----	--------	------------	-----	----
concluding	that	George	Washington		set the	precedent.	It
oneexample)encouragesprematurestoppingandweakens
thenformulatesafollow-upquerytoretrievehisinaugura-
								multi-turnreasoning.		Incontrast,thesimpler320curricu-
---	---	---	---	---	---	---	---	--------------------	---	----------------------------------	---	---	---	---
tiondateandsuccessfullyextractsthecorrectanswer,April
lummaintainsstrongerperformancebyallowingthemodel
30,1789. ThiscaseillustratesICRL’sabilitytodecompose
toexplorelongerreasoningpathsduringtraining.
complexquestions,retrieverelevantinformationacrossmul-
tipleturns,andmaintaincoherentreasoningwithoutexplicit
intermediatesupervision. Itdemonstratestheeffectiveness ModelScalingand14BPerformance. Toevaluatethe
ofourframeworkinlearningstructuredtool-usebehaviors scalabilityofICRLacrosslargermodels, weappliedour
7

In-ContextReinforcementLearningforToolUseinLargeLanguageModels

	(a)ResponseLength			(b)Reward		(c)NumberofValidSearch
Figure3.Trainingdynamicscomparisonacrossdifferentfew-shotsettings(3-shot,2-shot,0-shot)forQwen-7Bmodel.
Table7.Results:Accuracy(%)onMathQAdatasets.
methodtoQwen2.5-14B-InstructandreportresultsinTa-
ble6.ICRLsignificantlyoutperformsbothdirectprompting
MathQA
andCoTmethods acrossallfiveQAdatasets. Inparticu- Model Method SFT
							AIME2024	AIME2025
---	---	---	---	---	---	---	--------	--------	---
lar, itachieves75.0EMonTriviaQAand61.8on2Wiki,
					ReTool	✓	67.0		49.3
-----------------------------------------------	-------------------	---------	---------------	--------	------	---	----	---	----
yieldingastrongaverageEMscoreof51.84—surpassing				Qwen3-8B
					ICRL	✗	64.1		51.7
CoT	by +20.7 and direct	prompting	by +27.0. These	re-
sultsdemonstratethatICRLcontinuestoscaleeffectively
tolargermodelsizesandbenefitsfromincreasedcapacity,
compareourresultswithReTool,whichisaSFT-RLtrain-
withoutrequiringadditionalsupervisionorannotation.
ingframeworkthattrainingmodelstolearncode-writing
				and tool-calling	on running	code.	ReTool	(Feng	et al.,
---	---	---	---	----------------	----------	-----	------	-----	-------
4.2.Trainingprocess
				2025) achieves	the state-of-the-art		performance	on	code-
---	---	---	---	--------------	--------------------	---	-----------	---	-----
TounderstandhowICRLevolvesduringtraining,weana-
augmentedlong-formreasoningforsolvingmathproblems,
lyzethelearningcurvesacrossthe3-shot,2-shot,and0-shot butitalsoneedsalotofannotateddatatoapplycold-start
curriculumstagesusingtheQwen2.5-7Bmodel.Intheearly
SFTformodelstolearnthetool-callingformatandunder-
stageswithdemonstrations(3-shotand2-shot),themodel standthewholereasoningprocess. However,ourmethod
producesrelativelystableandwell-structuredresponses,as ICRLdoesn’tneedtoSFTthemodelfirst,andmodelscan
reflected	in the consistent	response	lengths. As training
---------	-----------------	--------	--------------------	---	---	---	---	---	---
stillreasoningproperlyfromtheIn-ContextLearningwith
progressesintothe0-shotstage,theresponselengthinitially the examples we provides in prompt. From Table 7, al-
dropsduetotheremovalofin-contextexamplesbutgradu-
throughoutmethodunderperformsReToolonAIME2024
allyincreasesagain,indicatingthatthemodelislearningto by 2.9%, it can achieve better result on AIME2025 with
independentlycomposelongerandmorestructuredoutputs.
				+2.4%accuracy.	Whichmeansourmethodstillworksfor
---	---	---	---	--------------	--------------------------------	------------------	----------	---	------
				help models	to learn	other tool-calling	operations		and is
Althoughtherewardremainsrelativelysteadythroughout
moredata-efficientthanothermethodsthatneedcold-start
trainingandisbasedonlyonsparsesignals—outputformat
validityandfinalansweraccuracy—themodelstilllearns SFTwiththousandsofannotateddata.
tousetoolsmoreeffectivelyovertime.			Thisismostclearly
----------------------------------------------------	------------------------------------	---	-----------------	------------	---	---	---	---	---
reflectedintheincreasingnumberofvalidtoolcallsduring				5.Conclusion
the0-shotphase.	Theriseinvalidtoolusageindicatesthat
WeintroduceICRL,asimpleyetpowerfulframeworkfor
ICRLsuccessfullyencouragesthemodeltointernalizetool-
usebehavior,evenwithoutdenseorstep-levelsupervision. training LLMs to use tools via in-context reinforcement
				learning,withoutrequiringSFTorlabeledtooltraces.					By
---	---	---	---	------------------------------------------------	---	---	---	---	---
incorporatingfew-shotdemonstrationsdirectlyintotheRL
4.3.Generalize
rolloutpromptsandgraduallyphasingthemout,ICRLen-
Apart from conducting our method in web search tool- ablesmodelstotransitionfromimitationtoautonomoustool
callingdomain,wealsoevaluateourmethodbytrainingthe
				usethroughreward-drivenlearning.			Ourmethodachieves
---	---	---	---	--------------------------------	---	---	-----------------	---	---
models’abilitiesoncode-writingandcallingtoolstorunthe strong performance across a range of QA and reasoning
pythoncodetohelpthemsolvecomplexmathproblems.We benchmarks, outperforming existing approaches that rely
8

In-ContextReinforcementLearningforToolUseinLargeLanguageModels onsuperviseddataorfrozentool-usepolicies. ICRLalso Hsiao, V., Fine-Morris, M., Roberts, M., Smith, L. N., generalizesacrossdomains,includingwebsearchandcode and Hiatt, L. M. A critical assessment of LLMs for execution, demonstrating its flexibility and effectiveness. solving multi-step problems: Preliminary results. In TheseresultshighlightICRLasascalableanddata-efficient AAAI 2025 Workshop LM4Plan, 2025. URL https: alternative to traditional SFT+RL pipelines for enabling //openreview.net/forum?id=kFrqoVtMIy. tool-augmentedlanguagemodels.

				Jin, B., Zeng,	H.,	Yue, Z.,	Yoon,	J.,	Arik, S.,	Wang, D.,
				Zamani,	H., and	Han,	J. Search-r1:		Training	llms to
ImpactStatements
reasonandleveragesearchengineswithreinforcement
Thispaperpresentsworkwhosegoalistoadvancethefield learning. arXivpreprintarXiv:2503.09516,2025a.
of machine	learning. There	are many potential	societal
----------	---------------	------------------	--------	---	---	---	---	---	---	---
consequencesofourwork,noneofwhichwefeelmustbe Jin, J., Zhu, Y., Dou, Z., Dong, G., Yang, X., Zhang, C.,
				Zhao, T.,	Yang,	Z., and	Wen,	J.-R.	Flashrag:	A mod-
---	---	---	---	---------	-----	-------	----	-----	---------	------
specificallyhighlightedhere.
ulartoolkitforefficientretrieval-augmentedgeneration
				research.	InCompanionProceedingsoftheACMonWeb
---	---	---	---	---------	-----------------------------------	---	---	---	---	---
References
Conference2025,pp.737–740,2025b.
Ahn,J.,Verma,R.,Lou,R.,Liu,D.,Zhang,R.,andYin,W.
Largelanguagemodelsformathematicalreasoning: Pro- Joshi, M., Choi, E., Weld, D. S., and Zettlemoyer, L.
				Triviaqa:	A large	scale	distantly	supervised		challenge
---------------------	---	------------------------------	---	---------	-----------	--------------	---------	----------	-----	---------
gressesandchallenges.		arXivpreprintarXiv:2402.00157,
2024.				dataset	for reading	comprehension.			arXiv	preprint
arXiv:1705.03551,2017.
Cheng,J.,Marone,M.,Weller,O.,Lawrie,D.,Khashabi,
D., and Van Durme, B. Dated data: Tracing knowl- Kwiatkowski, T., Palomaki, J., Redfield, O., Collins, M.,
edge cutoffs in large language models. arXiv preprint Parikh,A.,Alberti,C.,Epstein,D.,Polosukhin,I.,Kel-
arXiv:2403.12958,2024. cey, M., Devlin, J., Lee, K., Toutanova, K. N., Jones,
				L., Chang,	M.-W.,	Dai,	A., Uszkoreit,		J.,	Le, Q., and
---	---	---	---	----------	------	----	--------------	---	---	-----------
Chung,H.W.,Hou,L.,Longpre,S.,Zoph,B.,Tay,Y.,Fe- Petrov,S. Naturalquestions: abenchmarkforquestion
dus,W.,Li,Y.,Wang,X.,Dehghani,M.,Brahma,S.,etal. answeringresearch. TransactionsoftheAssociationof
Scalinginstruction-finetunedlanguagemodels. Journal ComputationalLinguistics,2019.
ofMachineLearningResearch,25(70):1–53,2024.
Lewis,P.,Perez,E.,Piktus,A.,Petroni,F.,Karpukhin,V.,
Feng,J.,Huang,S.,Qu,X.,Zhang,G.,Qin,Y.,Zhong,B., Goyal,N.,Küttler,H.,Lewis,M.,Yih,W.-t.,Rocktäschel,
Jiang,C.,Chi,J.,andZhong,W. Retool: Reinforcement T.,etal. Retrieval-augmentedgenerationforknowledge-
learning for strategic tool use in llms. arXiv preprint intensivenlptasks. Advancesinneuralinformationpro-
arXiv:2504.11536,2025.				cessingsystems,33:9459–9474,2020.
----------------------	---	---	---	---------------------------------	---	---	---	---	---	---
Gao,Y.,Xiong,Y.,Gao,X.,Jia,K.,Pan,J.,Bi,Y.,Dai,Y., Li,X.,Dong,G.,Jin,J.,Zhang,Y.,Zhou,Y.,Zhu,Y.,Zhang,
Sun, J., Wang, H., andWang, H. Retrieval-augmented P., and Dou, Z. Search-o1: Agentic search-enhanced
generationforlargelanguagemodels: Asurvey. arXiv largereasoningmodels.arXivpreprintarXiv:2501.05366,
preprintarXiv:2312.10997,2(1),2023.				2025a.
-----------------------------------	---	---	---	------	---	---	---	---	---	---
Guo,D.,Yang,D.,Zhang,H.,Song,J.,Zhang,R.,Xu,R., Li,X.,Jin,J.,Dong,G.,Qian,H.,Zhu,Y.,Wu,Y.,Wen,J.-
Zhu,Q.,Ma,S.,Wang,P.,Bi,X.,etal. Deepseek-r1: In- R.,andDou,Z.Webthinker:Empoweringlargereasoning
				models	with deep	research	capability.		arXiv	preprint
---	---	---	---	------	---------	--------	-----------	---	-----	--------
centivizingreasoningcapabilityinllmsviareinforcement
learning. arXivpreprintarXiv:2501.12948,2025. arXiv:2504.21776,2025b.
				Matarazzo,A.andTorlone,R.			Asurveyonlargelanguage
--------	---------------	--------------	----------------	-------------------------	---	---	----------------------	---	---	---
Guo, T.,	Chen, X., Wang,	Y., Chang, R.,	Pei, S., Chawla,
N.V.,Wiest,O.,andZhang,X. Largelanguagemodel modelswithsomeinsightsontheircapabilitiesandlimi-
basedmulti-agents: Asurveyofprogressandchallenges. tations. arXivpreprintarXiv:2501.04040,2025.
InIJCAI,2024.
				Mei, J., Hu,	T.,	Fu, D.,	Wen,	L.,	Yang,	X., Wu, R.,
---	---	---	---	------------	---	-------	----	---	-----	-----------
Ho, X., Nguyen, A.-K. D., Sugawara, S., and Aizawa, Cai, P., Cai, X., Gao, X., Yang, Y., et al. O2-
A. Constructing a multi-hop qa dataset for compre- searcher: A searching-based agent model for open-
hensive evaluation of reasoning steps. arXiv preprint domainopen-endedquestionanswering. arXivpreprint
arXiv:2011.01060,2020.				arXiv:2505.16582,2025.
----------------------	---	---	---	----------------------	---	---	---	---	---	---
9

In-ContextReinforcementLearningforToolUseinLargeLanguageModels Nguyen,X.-P.,Pandit,S.,Reddy,R.G.,Xu,A.,Savarese, for knowledge-intensive multi-step questions. In Pro- S., Xiong, C., and Joty, S. Sfr-deepresearch: Towards ceedingsofthe61stannualmeetingoftheassociationfor effective reinforcement learning for autonomously rea- computational linguistics (volume 1: long papers), pp.

soningsingleagents.		arXivpreprintarXiv:2509.06283,					10014–10037,2023.
2025.
							Wang, C.,	Deng, Y.,	Lyu,	Z., Zeng, L.,	He, J., Yan,	S.,
---	---	---	---	---	---	---	---------	---------	---------	-------------	------------	---
							and An,	B. Q*:	Improving	multi-step	reasoning	for
Press,O.,Zhang,M.,Min,S.,Schmidt,L.,Smith,N.A.,
andLewis, M. Measuringandnarrowingthecomposi- llms with deliberative planning, 2024a. URL https:
tionalitygapinlanguagemodels. InFindingsoftheAs- //arxiv.org/abs/2406.14283.
sociationforComputationalLinguistics:				EMNLP2023,
-------------------------------------	---	---	---	----------	---	---	---	---	---	---	---	---
Wang,S.,Zhu,Y.,Liu,H.,Zheng,Z.,Chen,C.,andLi,J.
pp.5687–5711,2023.
							Knowledgeeditingforlargelanguagemodels:				Asurvey.
---	---	---	---	---	---	---	---------------------------------------	---	---	---	--------	---
Qu, C., Dai, S., Wei, X., Cai, H., Wang, S., Yin, D., ACMComputingSurveys,57(3):1–37,2024b.
Xu, J.,	and Wen,	J.-r.	Tool learning	with	large	lan-
-------	--------	-----	-------------	----	-----	----	---	---	---	---	---	---
Wei,J.,Wang,X.,Schuurmans,D.,Bosma,M.,Xia,F.,Chi,
guage models:		a survey.	Frontiers	of Computer		Sci-
-------------	---	---------	---------	-----------	---	----	---	---	---	---	---	---
ence, 19(8), January 2025. ISSN 2095-2236. doi: E.,Le,Q.V.,Zhou,D.,etal.Chain-of-thoughtprompting
							elicitsreasoninginlargelanguagemodels.				Advancesin
---------------------------	---	---	-----------------	---	---	---	--------------------------------------	---	---	---	----------	---
10.1007/s11704-024-40678-2.			URLhttp://dx.doi.
neuralinformationprocessingsystems,35:24824–24837,
org/10.1007/s11704-024-40678-2.
2022.
Seed,B.,Chen,J.,Fan,T.,Liu,X.,Liu,L.,Lin,Z.,Wang,
Yang,A.,Yang,B.,Zhang,B.,Hui,B.,Zheng,B.,Yu,B.,
M.,Wang,C.,Wei,X.,Xu,W.,etal.				Seed1.5-thinking:
-----------------------------	---	---	---	-----------------	---	---	---	---	---	---	---	---
Li,C.,Liu,D.,Huang,F.,Wei,H.,Lin,H.,Yang,J.,Tu,
Advancingsuperbreasoningmodelswithreinforcement
J.,Zhang,J.,Yang,J.,Yang,J.,Zhou,J.,Lin,J.,Dang,
learning.	arXivpreprintarXiv:2504.13914,2025.
---------	-----------------------------------	---	---	---	---	---	---	---	---	---	---	---
K.,Lu,K.,Bao,K.,Yang,K.,Yu,L.,Li,M.,Xue,M.,
Zhang,P.,Zhu,Q.,Men,R.,Lin,R.,Li,T.,Tang,T.,Xia,
Shao,Z.,Wang,P.,Zhu,Q.,Xu,R.,Song,J.,Bi,X.,Zhang,
H.,Zhang,M.,Li,Y.,Wu,Y.,etal. Deepseekmath: Push- T., Ren, X., Ren, X., Fan, Y., Su, Y., Zhang, Y., Wan,
ingthelimitsofmathematicalreasoninginopenlanguage Y., Liu, Y., Cui, Z., Zhang, Z., and Qiu, Z. Qwen2.5
models. arXivpreprintarXiv:2402.03300,2024. technicalreport,2024a.
Sheng,G.,Zhang,C.,Ye,Z.,Wu,X.,Zhang,W.,Zhang, Yang,A.,Yang,B.,Zhang,B.,Hui,B.,Zheng,B.,Yu,B.,
R.,Peng,Y.,Lin,H.,andWu,C. Hybridflow: Aflexi- Li,C.,Liu,D.,Huang,F.,Wei,H.,Lin,H.,Yang,J.,Tu,
J.,Zhang,J.,Yang,J.,Yang,J.,Zhou,J.,Lin,J.,Dang,
bleandefficientrlhfframework.				arXivpreprintarXiv:
-----------------------------	---	---	---	-------------------	---	---	-------------------------------------------	---	---	---	---	---
2409.19256,2024.							K.,Lu,K.,Bao,K.,Yang,K.,Yu,L.,Li,M.,Xue,M.,
Zhang,P.,Zhu,Q.,Men,R.,Lin,R.,Li,T.,Xia,T.,Ren,
Shi, Z., Gao,	S.,	Yan, L.,	Feng, Y.,	Chen,	X., Chen,	Z.,
-------------	---	--------	---------	-----	---------	---	---	---	---	---	---	---
X.,Ren,X.,Fan,Y.,Su,Y.,Zhang,Y.,Wan,Y.,Liu,Y.,
Yin, D., Verberne, S., and Ren, Z. Tool learning in Cui,Z.,Zhang,Z.,andQiu,Z. Qwen2.5technicalreport.
the wild: Empowering language models as automatic arXivpreprintarXiv:2412.15115,2024b.
toolagents,2025.URLhttps://arxiv.org/abs/
2405.16533.							Yang,A.,Li,A.,Yang,B.,Zhang,B.,Hui,B.,Zheng,B.,
-----------	---	---	---	---	---	---	-----------------------------------------------	---	---	---	--------------	---
							Yu,B.,Gao,C.,Huang,C.,Lv,C.,etal.				Qwen3technical
Sun,H.,Qiao,Z.,Guo,J.,Fan,X.,Hou,Y.,Jiang,Y.,Xie, report. arXivpreprintarXiv:2505.09388,2025.
P., Zhang,	Y.,	Huang, F.,	and Zhou,	J. Zerosearch:		In-
----------	---	----------	---------	--------------	---	---	---	---	---	---	---	---
Yang,Z.,Qi,P.,Zhang,S.,Bengio,Y.,Cohen,W.,Salakhut-
centivizethesearchcapabilityofllmswithoutsearching.
arXivpreprintarXiv:2505.04588,2025. dinov,R.,andManning,C.D. Hotpotqa: Adatasetfor
							diverse, explainable		multi-hop	question	answering.	In
---------	--------	----	---------	---------	---	-----	--------------------	---	---------	--------	----------	---
Team, K.,	Bai, Y.,	Bao,	Y., Chen,	G., Chen,	J.,	Chen,
Proceedingsofthe2018conferenceonempiricalmethods
N., Chen, R., Chen, Y., Chen, Y., Chen, Y., et al. innaturallanguageprocessing,pp.2369–2380,2018.
Kimi k2:	Open	agentic	intelligence.	arXiv	preprint
--------	----	-------	-------------	-----	--------	---	---	---	---	---	---	---
Zhao,S.,Yu,T.,Xu,A.,Singh,J.,Shukla,A.,andAkkiraju,
arXiv:2507.20534,2025.
							R. Parallelsearch:		Trainyourllmstodecomposequery
---	---	---	---	---	---	---	------------------	---	-----------------------------	---	---	---
Trivedi,H.,Balasubramanian,N.,Khot,T.,andSabharwal, and search sub-queries in parallel with reinforcement
A. Musique: Multihop questions via single-hop ques- learning. arXivpreprintarXiv:2508.09303,2025.
		Transactions		of the Association		for
---	---	------------	---	------------------	---	---	---	---	---	---	---	---
tion composition.
Zhu,Z.,Liao,Y.,Chen,Z.,Wang,Y.,Guan,Y.,Wang,Y.,
ComputationalLinguistics,10:539–554,2022.
andWang,Y.Evolvebench:Acomprehensivebenchmark
Trivedi,H.,Balasubramanian,N.,Khot,T.,andSabharwal, for assessing temporal awareness in llms on evolving
A. Interleavingretrievalwithchain-of-thoughtreasoning knowledge. InProceedingsofthe63rdAnnualMeeting
10

In-ContextReinforcementLearningforToolUseinLargeLanguageModels oftheAssociationforComputationalLinguistics(Volume 1: LongPapers),pp.16173–16188,2025. 11