Datasets:

luoojason
/

muddle-eval-bundle-md

	-Bench:		Graduate-level			Multi-disciplinary			Benchmarks			for
		LLM	& MLLM		Complex		Reasoning		Evaluation
Meng-HaoGuo1 JiajunXu1 YiZhang1 JiaxiSong1 HaoyangPeng1 Yi-XuanDeng1 XinzhiDong1
KiyohiroNakayama2 ZhengyangGeng3 ChenWang4 BolinNi5 Guo-WeiYang6
YongmingRao†5 HouwenPeng†5 HanHu5 GordonWetzstein2 Shi-MinHu†(cid:66)1
5202 yaM 4 ]VC.sc[ 1v81020.5052:viXra Abstract
Reasoningstandsasacornerstoneofintelligence, 92.3 90.0 M M L U - B e n c h - T
								88.0				M M M U	- B e n c h - M
---	---	---	---	---	---	---	---	----	---	---	---	-------	---------------
enablingthesynthesisofexistingknowledgeto
78.2
solve complex		problems.	Despite	remarkable
-------------	---	---------	-------	----------	---	---	---	----	---	---	---	----	---
								69.0				69.1
progress,existingreasoningbenchmarksoftenfail
61.2
)%( ycaruccA
torigorouslyevaluatethenuancedreasoningcapa-									53.6			53.2
--------------------------------------------	---	---	---	---	---	---	---	---	----	---	---	----	---
bilitiesrequiredforcomplex,real-worldproblem-
solving,	particularly		in multi-disciplinary		and
--------	------------	---	---------------------	---	---	---	---	---	---	---	---	---	---
33.4
multimodal	contexts.		In this paper,		we intro-
----------	---------	---	--------------	---	---------	---	---	---	---	---	---	---	---
duceagraduate-level,multi-disciplinary,English-
Chinesebenchmark,dubbedasReasoningBench
(R-Bench),	for	assessing	the reasoning		capabil-
----------	---	---------	-------------	---	--------	---	---	---	---	---	---	---	---
R-
ityofbothlanguageandmultimodalmodels. o1-20241217 GPT-4o DeepSeek-R1 o1-20241217 GPT-4o
Benchspans1,094questionsacross108subjects
forlanguagemodelevaluationand665questions
							Figure1.Top-1	accuracy	comparison		of different	models	on
---	---	---	---	---	---	---	-------------	--------	----------	---	------------	------	---
across 83 subjects for multimodal model test- MMLU,MMMU,andR-Bench. R-Benchposesagreaterchal-
ing in both	English		and Chinese.	These	ques-		lengetocurrentmodels.
-----------	-------	---	------------	-----	-----	---	---------------------	---	---	---	---	---	---
tionsaremeticulouslycuratedtoensurerigorous
difficultycalibration,subjectbalance,andcross-
linguisticalignment,enablingtheassessmentto Reasoning,thesystematicprocessofsynthesizingknowl-
be an Olympiad-level multi-disciplinary bench- edgetosolvenovelproblems,liesattheheartofintelligence.
mark. Weevaluatewidelyusedmodels,includ- Yet,asfoundationmodelsgrowincreasinglysophisticated,
ingOpenAIo1,GPT-4o,DeepSeek-R1,etc.						Ex-
-----------------------------------	---	---	---	---	---	---	-------------------	---	----	------------------	---	------	-----
							existing benchmarks		fail	to comprehensively		assess	their
perimentalresultsindicatethatadvancedmodels complex reasoning capabilities. As shown in the above
performpoorlyoncomplexreasoning,especially quote,beforeequippingfoundationmodelswithreasoning
multimodalreasoning. Eventhetop-performing skills,weshouldfirstdefinegoalsforthembyestablishing
modelOpenAIo1achievesonly53.2%accuracy areliableevaluationtoassesstheirreasoningcapabilities.
onourmultimodalevaluation.			Dataandcodeare
--------------------------	---	---	--------------	---	---	---	---------------------------	---	---	---	-------------------	---	---
							Asnotedin(Kahneman,2011)and				(Weietal.,2022),re-
madepubliclyavailableathere.
							alizing system-I,		a.k.a., quick	and	intuitive	thinking	and
---	---	---	---	---	---	---	-----------------	---	-------------	---	---------	--------	---
system-II,a.k.a.,slowanddeliberatereasoningraisesdis-
							tinctrequirementsonfoundationmodels.					Similarly,assess-
---	---	---	---	---	---	---	------------------------------------	---	---	---	---	-----------------	---
1.Introduction
							ing quick	thinking	and complex		reasoning	requires	sub-
---	---	---	---	---	---	---	---------	--------	-----------	---	---------	--------	----
“Settinggoalsisthefirststepinturningtheinvisi- stantiallydifferentassessmentmethods. Ontheonehand,
evaluatingsystem-Ineedstoevaluatetheknowledgeand
bleintothevisible.”			—TonyRobbins
-------------------	---	---	------------	---	---	---	-------------	---	--------	----------	-------	-----	-------
							memory, which		requires	collecting	various	daily	conver-
*Secondauthorslistedrandomly,†Jointprojectlead,1Tsinghua
							sations and	knowledge-based		questions		e.g., concept	and
---------	---	---	---------	---	---	---	-----------	---------------	---	---------	---	-------------	---
2Stanford			3Carnegie
University, University, Mellon University, common sense questions. On the other hand, evaluating
4UniversityofPennsylvania,5TencentHunyuanX,6Fitten.Corre-
							system-IIrequiresevaluatingcomplexreasoningskills.						It
---	---	---	---	---	---	---	--------------------------------------------------	---	---	---	---	---	---
spondenceto:Shi-MinHushimin@tsinghua.edu.cn.
requiresgatheringadiverserangeofreasoningquestions,
suchasanalyticalanddeductiveones,whichismorechal-
1

R-Bench

						benchmarksasexamples.		MMLU(Hendrycksetal.,2021)
Table1.Comp.	denotes	comprehensiveness.		o1	saturation rep-
isacomprehensivebenchmarkformulti-disciplineunder-
resentso1(OpenAI,2024b)performanceonthisbenchmark.					It
--------------------------------------------------	---	---	---	---	---	---	---	---	---	---
standing,whichhasservedasacriticalguideforthedevelop-
reflectsthechallengethatthebenchmarkposestoadvancedmod-
mentoffoundationmodelsinrecentyears.However,consid-
els.
eringthecurrentlevelofmodelintelligence,thisbenchmark
Name		Comp.	o1Saturation		Language
----	---	-----	------------	---	--------	---	---	---	---	---
isclosetosaturation(e.g.,o1(OpenAI,2024b)hasachieved
		✓				92.3%accuracyonit).		Besides,itdosenottakemultimodal-
----	---	---	-----	---	---	-------------------	---	--------------------------------	---	---
MMLU			0.923		en
AIME@2024 ✗ 0.744 en ityandmultilingualismintoconsideration,whichisalsocrit-
R-Bench-T ✓ 0.690 en&zh icalforanidealreasoningtest. MMMU(Yueetal.,2024a)
						isaholisticevaluationformultimodalreasoningtests.				With
---	---	---	---	---	---	-------------------------------------------------	---	---	---	----
✓
MMMU			0.782		en
----	---	---	-----	---	---	---	---	---	---	---
thelaunchofo1(OpenAI,2024b),thisbenchmarkisalso
R-Bench-M		✓	0.532		en&zh
---------	---	---	-----	---	-----	-----------------------------------------	---	---------------------------------	---	------
						closetosaturation.		Also,itcannotbeusedtoevaluatelan-
						guagemodelsandignoresmultilingualtesting.				Weshow
thecomparisonofMMLU,MMMU,andR-BenchinFig.1
lengingtocollectandfilterthantheformer.Inthispaper,we
						andTab.1.	Frontiermath(Glazeretal.,2024)collectssome
---	---	---	---	---	---	---------	------------------------------------------	---	---	---
focusonbuildingareliablecomplexreasoningbenchmark
						challenging	problems	specifically	designed	for advanced
--------------	--------	------	------	---	----------	------------	---------	------------	--------	--------------
for both large	language	models	(LLMs)	and	multimodal
						mathematical	reasoning	evaluation,	which	indicates that
largelanguagemodels(MLLMs).
currentmodelsstillexhibitweaknessesinmathematicalrea-
Howcanwedesignanidealassessmentforcomplexreason- soning. However,itfallsshortincomprehensivenessand
ing? Webelievefollowingfourpropertiesarecritical. multilingualtesting. ThisalsoappliestoOmni-Math(Gao
						et al., 2024)	and AIME	(OpenAI,	2024b),	both of which
---	---	---	---	---	---	-------------	--------	--------	-------	-------------
• Comprehensiveness. Evaluating the intelligence of serveasbenchmarksfocusedonemployingmathematical
olympiadchallenges.
foundationmodelsisakintoevaluatinghumanintel-
ligence.	Wecannotfocusonjustoneaspect,suchas
--------	-----------------------------------	---	---	---	---	---	---	---	---	---
Inthispaper,ourgoalistobuildabenchmarkR-Benchthat
mathematics. Acomprehensiveevaluationisessential. alignswiththefourpropertiesweproposedforevaluating
						the reasoning	abilities	of intelligent	models.	To achieve
-------------	---	------------	----------	------	-------	-------------	---------	--------------	-------	----------
• Difficulty.		A meaningful	evaluation	should	exhibit
that,wefollowmorethan100collegecoursesfrom19de-
thecapabilitytoeffectivelydiscriminatebetweenthe
performanceofdifferentmodelsandprovidevaluable partmentsatTsinghuaUniversityandcollectchallenging
problemsfromtheirexams,textbooks,quizzes,homework,
insightsforguidingmodelimprovement.					Atpresent,
-----------------------------------	---	---	---	---	----------	---	---	---	---	---
etc. Aftermultipleroundsofrigorousscreeningbyexperts
foundationmodelsaredevelopingrapidly,andsome
andmodels,wefinallyselect1,094questionsspanning108
simplebenchmarkshavebeensaturatedandcannotpro-
subjectsforlanguagemodelsreasoningtest,and665ques-
videguidanceanddiscriminationforadvancedmodels.
tionscovering83subjectsformultimodalmodelsreasoning
• Multimodality. Weliveinamultimodalworld,con- test. Wewillpresentthedetailedscreeningprocessinthe
stantlyprocessingvariousvisualandlinguisticsignals. Sec2. AfterbuildingtheR-Benchbenchmark,wetestthe
Therefore,anidealbenchmarkshouldbedesignedto reasoningcapabilitiesofvariouspowerfulproprietarymod-
assessbothLLMsandMLLMs. elssuchaso1(OpenAI,2024b),GPT-4o(OpenAI,2024a),
Gemini(Teametal.,2023),Claude(Anthropic,2024a),and
• Multilingualism.		We	believe	that performing	com-
------------------	---	---	-------	---------------	----	---	---	---	---	---
open-sourcedmodelssuchasLlama3(Touvronetal.,2023),
plexreasoningismorechallengingthanunderstanding Qwen2.5(Yangetal.,2024),etc. Fromexperiments,our
multiplelanguages.		Amodelwithrobustcomplexrea-
------------------	---	---------------------------	---	---	---	---	---	---	---	---
observationsandfindingsaresummarizedasfollows:
soningskillsshouldbecapableofsolvingreasoning
problemsacrossdifferentlanguages.				Thisislikefora
---------------------------------	---	---	---	--------------	---	---	---	---	---	---
• Withtheemergenceofadvancedmodelslikeo1,exist-
humanexpert,heorshewillnotlosetheabilitytoad-
ingmultidisciplinaryevaluationshavenearlyreached
dressproblemsduetolanguagechanges.					Thus,assess-
----------------------------------	---	---	---	---	------------	-----------	------------------------------------	---	---	---
						saturation.	Besides,solelyrelyingonmathproblems,
ingmodelperformanceonequallydifficultquestions
e.g.,mathematicalolympiadproblems,maybringbias
across	languages	is essential.	It	will provide	insight
------	---------	-------------	---	------------	-------	------------------	---	---------------------------	---	---
						inmodelevaluation.		Therefore,thecommunityneeds
intowhetherthemodelhasgenuinelylearnedtoreason
						challenging		multi-disciplinary	benchmarks	to guide
---	---	---	---	---	---	-----------	---	------------------	----------	--------
orismerelyoverfittingtoaspecificlanguage.
foundationalmodelsinenhancingtheirreasoningabil-
ities,andthegoalofR-Benchistoaddressit.
Whiletherehavebeenattemptstocreateanidealreasoning
benchmark,tothebestofourknowledge,existingbench- • We illustrate from three dimensions — expert scor-
marks cannot incorporate all four of these key properties ing,modelscoring,andmodelthinkingtime—that
simultaneously. Here,wetakesomewidelyusedreasoning R-Bench is a more complex benchmark with higher
2

R-Bench requirementsformodelreasoningcomparedtoexisting multidisciplinarybenchmarksMMLUandMMMU. Step1: Step2:Experts collectand

• Multimodalcomplexreasoningremainschallenging.							Define a list
selectreasoningquestions
Despiterapidadvances,modelslagbehindtext-based							of collected
----------------------------------------------	---	---	---	---	---	---	-------------	---	---	---	---
KQ
disciplines
reasoning.	Forinstance,GPT-4oscores53.6%ontext
----------	-----------------------------------	---	---	---	---	---	---	---	---	---	---
RQ
butonly33.7%inmultimodalreasoningonR-Bench.
• ChainofThought(CoT)canenhancereasoningabili-						Step3:
----------------------------------------------	---	---	---	---	---	------	---	---	---	----	----
tiesinmostchatmodels,suchasGPT-4o.However,for										Some	Some
Digitize the questions
									text-only		multimodal
---	---	---	---	---	---	---	---	---	----------	---	----------
reasoningmodelslikeo1-mini,CoTdoesnothavethe
									questions		questions
-----------	---	-------------------------------	---	---	---	---	---	---	---------	---	---------
sameeffect.		Thismaybebecausereasoningmodels
inherentlybuildCoT,makingexplicitCoTineffective.
• ModelsmaintainhighconsistencyinansweringChi-
neseandEnglishquestionsofequaldifficulty,exceed- Step4:o1 model rescreen
						Step5:The third round			based on reasoning difficulty
---	---	---	---	---	---	----------------------	---	---	-----------------------------	---	---
ing70%formostmodels,demonstratingstrongcross-
screensfor completeness,
lingualreasoningcapabilities.										<2000
-----------------------------	---	---	---	---	---	-------------	---------	---	---	-----	---
						repetitionand	ambiguity
AQ
• Foundation	models	perform	differently	across	disci-
------------	----------	-------	----------------	-----------	------	---	---	---	---	---	-----
plines.	Specifical	ly , G	P T -4o achieves	30.4%–68.3%
		ℛ -B e	nc h- T					CQ			>2000
accuracyacrossvariousfields.
2.R-Bench
!-Bench
									Step6:	Constructing
---	---	---	---	---	---	---	---	---	------	-------------	---
optionsandtranslations
Inthissection,wewillthoroughlyintroducetheconstruc-
tionprocessofR-Bench.						ℛ-Bench		ℛ-Bench
---------------------	---	----------------------------	---	---	---	-------	---	-------	---	---	---
		Theentireprocessinvolvesmul-				-T		-T(zh)
tiplestepssuchasdatacollection,filteringandimproving.
						ℛ-Bench		ℛ-Bench
---	---	---	---	---	---	-------	---	-------	---	---	---
TheoverallpipelineisillustratedinFig.2.
						-M		-M(zh)
---	---	---	---	---	---	---	---	------	---	---	---
2.1.Datacollection
Before gathering reasoning questions, we conducted an Figure2.PipelineofbuildingR-Bench. Theprocessisdivided
intosixsteps,whicharedetailedinSec.2.Thefunnelrepresents
investigation	of the	curriculum	systems	of graduate	and
-------------	------	----------	-------	-----------	---	---	---	---	---	---	---
screening.WealwaysRfiBlteenrcohuttheblueballandpreservethebrown
undergraduatestudentsacross19diffeCrentdepDartmentsat
-TC
TsinghuaUniversity. Basedonoursurvey,weobtaineda one. InStep2,KQandRQdenoteknowledge-basedquestions
			B		E	andreasoning-basedquestions,respectively.				InStep4,<2000
---	---	---	---	---	---	-----------------------------------------	---	---	---	-------------	---
collectionlistcoveringover100coursesacross19depart-
indicatesthatthereasoningtokensofo1arelessthan2000.Finally,
ments,whichisshowninFig.2Step1.						RBench		RBench
-------------------------------	---	---	---	---	---	-------------	-------------------------------------------	------	---	---	---
			A			inStep5,-AMQE	andCQ-rMeCpresentambiguousquestionsandclear
F
After acquiring a collection list, we recruit senior under- questions,respectively. -Tindicatestext-onlytestingforLLMs.
graduatesandgraduatestudentsfromdifferentdepartments -Mmeansmultimodaltesting.zhrepresentstheChineseversion.
asexpertstoprovidereasoningquestion-answerpairs.					We
------------------------------------------------	---	---	---	---	---	---	---	---	---	---	---
recruitatotalof51experts,withatleasttwoparticipants
from each	department,	to	help us collect	and filter	ques-
----------------------------------	-------------	---	---------------	---------------	---------	----------	---------	------	----------------	-------	------------
						After the	two steps	above,	we collected	a total	of 10,270
tions. Duringthecollectionprocess,				wemainlyfocuson
						questions.	Among	them,	7,163 questions,		which do not
controlling	the following		key aspects:	1) The	questions
includeimages,aredesignatedfortestinglanguagemodels,
should align with the collection list we provide. 2) The whiletheremaining3,107questions,containingimages,are
professionalexpertiseshouldfilterout“knowledge-based”
allocatedfortestingmultimodalmodels.
questions—those		that rely	solely on memory	rather	than
---------------	----------------------------------	---------	----------------	------	--------	---	---	---	---	---	---
reasoning,	suchasconcept-definitionquestions.				Simulta-
2.2.Datadigitization
neously,	experts should	retain	reasoning-based		questions
--------	--------------	------	---------------	---	---------	---	---	---	---	---	---
andensuretheypresentasufficientdegreeofdifficulty. 3) Afterinitiallycollectingthequestions,wefindthatthecol-
Allquestionsshouldhavecorrespondinganswersthatcan lectedquestionsareinamessyformat,includingpictures,
beautomaticallyverified. Inthiscollectingprocess,weex- screenshots, text, etc. In addition, the summary question
cludeproof-basedquestions,ascurrentautomatedmethods filesprovidedbydifferentexpertsarealsodifferent,includ-
cannotverifythecorrectnessofproofs. Theprocessabove ingpdf,word,excel,etc. Therefore,weneedtoorganize
isshowninFig.2Step2.						anddigitizethisdata.
--------------------	---	---	---	---	---	--------------------	---	---	---	---	---
3

R-Bench Examples for language models reasoning ability test

		Major: computerscience;				Major:math;
		Subject:data structure.				Subject:complex variable function.
In the undirected graph G=(V,E) where
V={1,2,3,4,5,6,7} and E={(1,2), (1,3), (2,3), English Chinese
(4,5), (3,6), (4,7), (5,7)}, how many different Consider the polynomial p ( z ) = z 5 + z 3 + 5 z 2 + 2 . p(z)=z5+z3+5z2+2。
	s p a n n	i n g f o r e s ts d o e s th e gr	a p h G c o n t a in ?				考虑多项式
---	----------	------------------------------------------	-----------------------------	---------	-----------------------------	-----------------------------------	-----------------------------	---	---	---
				H o w m	a n y z e r o p o in t s (	c o u n ti n g m u ltiplicities)	在环域 1<	z	<2 中p 有多少个零点(计入重数)?
	N o t e :	I f t h e e d g e s e ts o f t w o	s p a nn in g f o re s ts	do e s p	h a v e i n th e a n n u	l u s 1 <	z	< 2 ?
are different, they are considered different
	spanning forests.			A:5 B:3			A:5 B:3
---	------------------	---	---	---------	---	---	---------	---	---	---
C:2 D:其他答案都不正确
C:2 D:All other answers are incorrect
	A:14 B:9 C:10 D:12 E:11			E:4 F :6			E:4 F :6
---	-------------------------------	---	---	----------	---	---	----------	---	---	---
F:All other answers are incorrect
Examples for multimodal models reasoning ability test
Major:mechanical engineering;
Subject:theoretical mechanics.
		English					Chinese
---	---	-------	---	---	---	---	-------	---	---	---
Thewidthofthebrickclampis25cm,and
	thecurvedrodsAGBandGCEDarehinged					砖夹的宽度为	25 cm , 曲杆 AGB
---	---------------------------------------	-----------------------------	---	---	---	-----------------	--------------	---	---	---
						与在	点铰接，尺寸如图
	atpointG,withdimensionsasshowninthe					GCED	G
	figure.Supposetheweightofthebrickis					所示。设砖重	Q=120N 提起砖的力
	Q=120NandtheforcePthatliftsthebrick					P 作用在砖夹的中心线上，砖夹与砖
						间的摩擦系数	f=0.5, 若想把砖夹起，
	actsalong	thecenterlineofthebrickclamp.
	Thecoefficientoffrictionbetweenthebrick					试求距离b的最大值？
clampandthebrickisf=0.5.Determinethe
A : 9 cm B : 11 cm C : 15 cm
maximumvalueofdistancebrequiredtolift
	thebrickusingtheclamp.					D:20 cm E : 25 cm
---	----------------------	---	---	---	---	------------------------	---	---	---	---
F: 其他答案都不正确
		A : 9 cm B : 11 cm	C : 15 cm
---	---	--------------------------	---------	---	---	---	---	---	---	---
D : 20 cm E : 25 cm F : All other answers are incorrect
Figure3.SomeexamplesinR-Bench.TheseexamplesshowthatR-Benchismultidisciplinary,multimodal,andmultilingual.Asshown
inthefigure,theproblemsinR-Bencharecomplexandcannotbesolvedbyquickthinking,whichshowsthatR-Benchfocusesondeep
reasoningproblemsratherthanknowledgeproblems,suchasconceptualproblems.
To	do so,	we recruit a data	annotation	team of	about 20	2.3.Datafiltering
-------	------	--------------------	---------------	-------	-----------	-----------------	---	---	---	---
people.		They are responsible	for organizing,		digitizing,
AsshowninFig.2,thefunnelsinsteps2,4,and5represent
checking,andcompilingallthequestionsintoExcelsheets.
threedifferentroundsofdatafiltering.Thesethreeroundsof
Thequestionsusedforlanguagemodelsareorganizedin
screeningrepresentexpertscreening,model-basedfiltering,
thefollowingformat:
andmanualreview.
“Department		- Subject -	Question (text)	- Answer	(text)	-
----------------------------------------------	---	-----------	---------------	--------	---------	---	---	---	---	---
OriginalQuestion(text,screenshots,photos,etc.)					-Original
Answer(text,screenshots,photos,etc.)”. Expert-screening. AsmentionedinSec.2.1,werecruit
expertsfromdifferentdepartmentstoprovidequestionsfor
Asforquestionsdesignedformultimodalmodels,theyare
						us. They primarily	rely on their	professional	knowledge
---	---	---	---	---	---	------------------	-------------	------------	---------	---
organizedintothefollowingformat:
tofilterout“knowledge-based”questionswhileretaining
“reasoning-based”questions.
“Department		- Subject -	Question (text)	- Answer	(text)	-
-----------	---	-----------	---------------	--------	------	---	---	---	---	---
QuestionImages-OriginalQuestion(text,screenshots,pho-
tosetc.) -OriginalAnswer(text,screenshots,photosetc.)”.
						Model-screening.	OpenAI	o1 (OpenAI,	2024b)	is a
---	---	---	---	---	---	----------------	------	-----------	------	----
Inthisprocess,weutilizetoolssuchasGPT-4oandMathpix widely used reasoning model. When we call its API, it
						returns the	number of reasoning	tokens,	which,	to some
---	---	---	---	---	---	-----------	-------------------	-------	------	-------
forOCRprocessing,followedbymanualproofreadingto
ensureitiscorrect. Afterthedatateamorganizesthedata, extent,reflectsthedifficultyofthequestion. Inthisround
weperformadouble-checkontheOCRresults. ofscreening,wemainlyfocusonthedifficultyofreasoning.
Wefilteroutthequestionswithlessthan2,000reasoning
tokenstoensurethatourR-Benchisabenchmarkforrea-
soningevaluation.
4

R-Bench Inorganic Chemistry Analytical Mechanics

						C h	e m i c a l T h e r m o d y n a m i c s						F u n d a m e n t a l P	h ysics
C h e m i s t	ry					P h	y s i c a l C h e m i s t r y
M a t h			3.7%3.7%3.4%										E l e c t r o m a g n e t	i s m
4.4% C h e m i c a l R e a c t i o n K i n e t i c s 4.1%3.9%3.8% E l e c t r o d y n a m i c s
P h y s ic s				2.9%		C o	m p u t a t i o n a l C h e m i s t r y
--------------	---	---	---	----	---	-----	-----------------------------------------	---	---	---	---	---	---	---	---
C o m p u t e r S. 4.7% 2.3% O rg a n i c C h e m i s t r y 4.7% 3.8% O p t i c s
E l e c t ro n	ic E.				2.3 %						3.5%
----------------	------	----	---	---	-----	------------------	----	---	----	---	----	----	---	---	---
		4.8%			2 .1%	Quantum Chemistry			4.8%
A e r o s pa	c e											3.2%
Automation					2 %
					1 .6%			5%				2.7%
Statistics		5.4%										2.6%
Mechanical E.							4.6%
Materials
								5.9%					3.8%
----------------	---	----	---	---	---	-----	----	----	---	---	---	---	-------	---	---
B i o l o g y		7.3%
C i v i l E .						14.7%	4.4%						9% 1.5%
1.5%
Vehicle E.								6.6%
-----------------	------	----	---	---	---	----	----	----	----	---	-----	---	----	---	---
Physics E.							2%						1.4%
C h e m i ca	l E.	8.1%
						1.8%	1%						0.9%
Ec o n o m	i c s										10.2%
Microelectronics							0.5%		7.5%
17.2%
Environment			9.4%				0.5%
-------------	---	------------------------	----	---	---	---	----	---	------------------------	---	----	---	---	---	---
Architecture									9.2%		9.8%
		(a)StatisticsofR-Bench-T							(b)StatisticsofR-Bench-M
Figure4.AccordingtostatisticsonR-Bench,thebenchmarkspans19departments,includingmathematics,physics,biology,computer
science,andchemistry,coveringover100subjectssuchasInorganicChemistry,ChemicalReactionKinetics,andElectromagnetism.It
features1,094questionsdesignedfortestinglanguagemodelsand665questionsspecificallytailoredforevaluatingmultimodalreasoning
capabilities.Foradetailedlistofsubjects,pleaserefertotheappendix.
2.5.OverviewofR-Bench
Manualreview.		Ourmanualreviewfocusesonwhether
-------------	---	-------------------------------	---	---	---	---	---	---	---	---	---	---	---	---	---
thequestionconditionsarecomplete,whetherthequestions
							After	completing	the	aforementioned		steps,	we develop
--------------------------------------------------	---	---	---	---	---	---	--------	----------	---------------	--------------	-----------------	------	------------	---	---
arerepeated,whetherthequestionsareambiguous,andthe							R-Bench,
								a	graduate-level,		multi-discipline,		multilingual
balanceofsubjects.
benchmarkdesignedtoevaluatecomplexreasoningcapa-
Checking for completeness, repetition, and ambiguity re- bilitiesforbothlanguageandmultimodalmodels. Fig.3
quire multiple rounds of thorough review by different in- illustrates several examples from R-Bench, clearly high-
dividuals to eliminate ambiguities, along with the use of lightingitsabovedistinctivefeatures.
duplicationdetectiontoolstoavoidrepetition.						Asforthe
-------------------------------------------	---	---------	-------	---------	-------	--------	-------	---	----------	---	---------	---------------	---	---	---
							R-Bench	can	be divided		into four	sub-benchmarks:		R-	1
balance check,		to reduce	testing	bias from	subject	imbal-
Bench-TandR-Bench-T(zh)forlanguagemodelevaluation,
ance, we	limit	the number	of questions		per	subject	to a
--------	-----	----------	------------	---	---	-------	---------	---	-------------	---	---	--------------	---	-----	---
							R-Bench-M		R-Bench-M(zh)
									and			for multimodal		model
maximumof50byfilteringoutexcess.
							evaluation.	Here,R-Bench-TdenotesR-Benchusingtext-
---	---	---	---	---	---	---	--------------	--------------------------------------	----------	---	-------	-----------	-------	---	---
							only questions		in English		for LLM	evaluation,	whereas
2.4.Conductingoptionsandtranslations
R-Bench-T(zh)representsR-Benchusingtext-onlyques-
							tionsinChineseforLLMevaluation.					Likewise,theother
--------	---------	-------	-----	---------	-------------	---	-------------------------------	---	---	---	---	-----------------	---	---	---
In order	to enable	answers	to be	evaluated	automatically
and accurately, we convert all questions such as analyti- twonotationsfollowthesamenamingconvention.
cal,fill-in-the-blank,andmultiple-choicequestionsintothe
WeconductstatisticalanalysisonR-Benchwiththeresults
single-choicequestionformat. WeuseGPT-4otoconstruct ItpresentstheR-Bench-Tstatisticsfor
presentedinFig.4.
5 options	for each	question	and	add an	option	“All	other
---------	--------	--------	---	------	------	----	-----	---	---	---	---	---	---	---	---
text-onlyquestionsusedinevaluatingthereasoningcapabil-
answersareincorrect”,whichequipseachquestionwith6 R-Bench-Tspans18departments,
itiesoflanguagemodels.
candidate	answers.	Then,	we check	the	options	multiple
---------------------------------------------	--------	-----	--------	---	-------	--------	---------------------	----------	------------	--------	----------	-------	--------------	---	---
							includingmathematics,			biology,	chemistry,		computersci-
timestoensurethecorrectnessofourconstruction.						Further-
							ence,	electronic	engineering,		and	others.	It encompasses
more,wemanuallyadjusttheoptionstoensureasufficient
over108subjects,suchascalculus,numbertheory,analytic
numericalgapbetweenthem,therebyavoidingerrorscaused
							geometry,	ordinary		differential	equations,		and functional
---	---	---	---	---	---	---	---------	--------	---	------------	----------	---	--------------	---	---
bynumericalapproximations. analysis,andcomprisesatotalof1,094questions. Fig.4
alsopresentsthestatisticsofR-Bench-M,whichevaluates
Besides,inordertoenableR-Benchtoacquirethemulti-
							thereasoningcapabilitiesofmultimodalmodels.						R-Bench-
---	---	---	---	---	---	---	-------------------------------------------	---	---	---	---	---	--------	---	---
lingualproperty,wemanuallyconstructedEnglish-Chinese
translationsforeachquestion. Duringthetranslationpro- Mincorporatesadiversesetofquestiontypesrequiringboth
cess,weutilizetoolslikeGPT-4o. Eachquestionismeticu- textualandvisualinputs. Itcovers18departments,suchas
physics,biology,architecture,andeconomics,andincludes
louslyreviewedandrefinedbythreeexpertsfluentinboth
EnglishandChinesetoensurecorrectnessandclarity. 83 subjects, such as thermodynamics, molecular biology,
structuraldesign,andmicroeconomics,includingatotalof
							665questions.		ItisworthnotingthatweprovideEnglish
---	---	---	---	---	---	---	-------------	---	-----------------------------------	---	---	---	---	---	---
andChineseversionsforallquestions.
5

R-Bench Table2.Comparisonofreasoningrequirementsforproblemsin Table4.Theaveragethinkingtimeofo1on30randomlyselected R-Bench-TandMMLUviaexpertando1voting. samplesfromdifferentbenchmarks.TTdenotesthinkingtime.

		R-Bench-Twin		MMLUwin	Tie		MMLU		R-Bench-T	MMMU		R-Bench-M
Expertvoting 85.94% 10.62% 3.44% TT 13.5s 98.2s(7.3×) 20.3s 91.7s(4.5×)
o1voting		76.67%		20.00%	3.33%
--------	---	------	---	------	-----	---	---	---	---	---	---	---	---
Table3.Comparisonofreasoningrequirementsforproblemsin
R-Bench-MandMMMUviaexpertando1voting.
3.2.Evaluatingreasoningcapabilityofdifferentmodels
R-Bench-Mwin
				MMMUwin	Tie
---	---	---	---	-------	---	---	---	---	---	---	---	---	---
WeemployR-Bench-Ttoassessthereasoningcapabilities
Expertvoting		76.88%		15.94%	7.19%
------------	---	------	---	------	-----	---	---	---	---	---	---	---	---
ofvariousLLMssuchaso1(OpenAI,2024b),GPT-4o(Ope-
o1voting		83.33%		13.33%	3.33%
--------	---	------	---	------	-----	---	---	---	---	---	---	---	---
nAI,2024a),DeepSeek-R1(AI,2025),Gemini(Teametal.,
2024),Claude3.5(Anthropic,2024b),Qwen2.5(Yangetal.,
						2024),	Llama3	(Dubey	et	al., 2024),	etc,	in both	English
---	---	---	---	---	---	-------------------	------	------	---------------------------------	-----------	----	-------	-------
						andChinesesettings.			TheevaluationinvolvesutilizingAPI
3.Experiments
						callsanddeployingopen-sourcemodelslocally.							ForAPI
---	---	---	---	---	---	------------------------------------------	---	---	---	---	---	---	------
calls,weutilizetheofficialinterfaceswithdefaulthyperpa-
AfterdevelopingR-Bench,weutilizeittoassessthecom-
rameters. Foropen-sourcemodels,wedeploytheirweights
plexreasoningcapabilitiesofvariousLLMsandMLLMs,
						locally	usingvLLM		(Kwon	et al.,2023),		settingthe	tem-
---	---	---	---	---	---	-------	---------	---	-----	-------------	---	----------	----
includingbothopen-sourcemodelssuchasLlamaandclose-
sourcemodelssuchasGPT-4o. Firstly,weaimtodemon- perature to 0 while keeping all other parameters at their
stratethatR-Benchisabenchmarkforcomplexreasoning default values. The evaluation was conducted using the
						toolsprovidedbyOpenCompass(Contributors,2023).							In
---	---	---	---	---	---	----------------------------------------------	---	---	---	---	---	---	---
throughexpertscoring(userstudy)ando1modelscoring.
						alltests,theCoTpromptisusedbydefault.						Fordetailson
---	---	---	---	---	---	-------------------------------------	---	---	---	---	---	------------	---
Then,weevaluatethereasoningcapabilitiesofmodelswith
						thespecificprompts,			pleaserefertoourappendix.				Inthe
---------------------------------------------	---	---	---	---	---	-------------------	---	---	-------------------------	---	---	---	-----
andwithoutCoTpromptingunderazero-shotsetting.					Fi-
resultsshowninTab.5,wefoundthatmodelsdesignedfor
nally,weanalyzetheexperimentsandsummarizeobserva-
						reasoning		tasks,	such as o1,	outperform		chat models	like
---	---	---	---	---	---	---------	---	------	-----------	----------	---	-----------	----
tionsandfindingsfromtheexperimentalprocess.
						GPT-4o	in	complex	reasoning.	Besides,		there	remains a
---	---	---	---	---	---	------	---	-------	----------	--------	---	-----	---------
significantgapincomplexreasoningbetweenopen-source
3.1.Reasoningcomparisonwithotherbenchmarks
modelsandcommercialmodels.
ToillustratethatR-Benchisabenchmarkdesignedtoeval-
R-Bench-M
						Moreover,		we utilize			to evaluate	the	reason-
--------------	--------	---	---------	------------	------	---------	---	----------	---	---	-----------	---	-------
uate reasoning	ability,		we employ	two methods:	expert
ingcapabilitiesofvariousMLLMs,includingo1(OpenAI,
scoringandreasoningmodelscoring.
2024b),GPT-4o(OpenAI,2024a),Claude3.5(Anthropic,
Weconductedexpertscoringthroughuserstudies.Tobespe- 2024b), Qwen2.5-VL (Yang et al., 2024), and InternVL
cific,werandomlyselected30questionsfromR-Bench-T 2.5(Chenetal.,2024),etc,acrossbothEnglishandChinese
andanother30questionsfromMMLUandpresentedthem languages. TheevaluationalsoinvolvesutilizingAPIcalls
to experts for pairwise comparisons to determine which anddeployingopen-sourcemodelslocally. ForAPIcalls,
questionrequiredmorereasoningskillstosolve. Wecon- weutilizetheofficialinterfaceswithdefaulthyperparam-
structed similar experiments using the same settings be- eters. For open-source models, we deploy their weights
tweenR-Bench-MandMMMU. locallyusingVLMEvalKit(Duanetal.,2024),settingthe
temperatureto0whilekeepingallotherparametersattheir
Forreasoningmodelscoring,weadoptedtwoapproaches.
						defaultvalues.		Inalltests,theCoTpromptisusedbyde-
------------	---	----	------	------------------	-----	--------------	---	----------------------------------	---	---	---	---	---
On one hand,	we	used	the o1	model to determine	which
fault. Fordetailsonthespecificprompts,pleaserefertoour
questionrequiredmorereasoningabilitybasedonthenum-
						appendix.		We draw	three	conclusions	from	the	results in
---	---	---	---	---	---	---------	---	-------	-----	-----------	----	---	----------
berofreasoningtokens(reasoningtime);ontheotherhand,
Tab.6. First,wefoundthatmodelsperformworseinmulti-
weaskedtheo1modeltodirectlycomparethetwoquestions
modalcomplexreasoningcomparedtoreasoninginapurely
anddeterminewhichonerequiredmorereasoning.
						linguistic		environment.	Second,		the reasoning		model o1
---	---	---	---	---	---	----------	---	------------	-------	---	-------------	---	--------
TheresultsareshowninTab.2,3and4. Theresultsindicate stilldemonstratesoutstandingperformanceinmultimodal
thatbotho1’sjudgmentandtheexperts’judgmentconsider complexreasoningevaluation. Third,thegapbetweenopen-
R-Bench to require significantly higher reasoning ability sourceandclosed-sourcemodelsisevenmorepronounced
comparedtoMMLUandMMMU.						inmultimodalcomplexreasoning.
----------------------	---	---	---	---	---	-----------------------------	---	---	---	---	---	---	---
6

R-Bench Table5.PerformancecomparisonofvariousmodelsonR-Bench- Table6.PerformancecomparisonofvariousmodelsonR-Bench- Tinzero-shotsettingswithCoT.Thetableisdividedbyamiddle Minzero-shotsettingswithCoT.Thetableisdividedbyamiddle line:API-basedmodelsarelistedabovetheline,whileopen-source line:API-basedmodelsarelistedabovetheline,whileopen-source modelsareshownbelow.‘zh’indicatestheChineseversion.The modelsareshownbelow.’zh’indicatestheChineseversion.The valuesinthetablerepresenttheTop-1accuracy,in%. valuesinthetablerepresenttheTop-1accuracy,in%.

	R-Bench-T	R-Bench-T(zh)			R-Bench-M	R-Bench-M(zh)
ModelName			ModelName
o1-20241217	69.0	70.1	o1-20241217		53.2	55.0
Gemini-2.0-flash-thinking 68.4 67.5 Claude3.5-sonnet@1022 39.7 38.3
Doubao1.5pro-20250121 62.0 63.4 Doubao1.5pro-20250121 37.9 42.4
o1-preview@20240912	62.3	62.6	GPT-4o-20241120		33.4	33.2
---------------------	----	----	----------------------	---	----	----
o1-mini@20240912	64.0	59.9	Gemini-1.5-Pro		35.5	35.9
Doubao-pro-20241215	60.7	60.8
			Qwen2-VL-72B		25.1	25.7
Claude3.5-sonnet@0620	57.5	57.0
			Qwen2-VL-7B		19.6	22.3
GPT-4o-20241120	53.6	51.6
			LLaVA-OneVision-7B		23.8	23.5
MiniMax-Text-01	53.8	53.6
			DeepSeek-VL2		21.8	24.4
GLM-Zero-Preview	53.6	48.6
			Llama3.2V-11B-Instruct		20.0	18.6
ERNIE-4.0-8K-Latest	39.7	50.1
			InternVL-2.5-8B		15.9	17.1
Deepseek-R1	61.2	59.3
Deepseek-V3	59.6	56.6
Qwen3-235B-A22B	58.0	58.4
Table7.AssessingtheperformanceimpactofCoTacrossdifferent
Qwen3-32B	52.3	54.3	modelsonR-Bench-T.
----------------------	----	----	---------------------	---	-------	---------
Qwen2.5-72B-Instruct	53.7	52.0
Llama-3.3-70B-Instruct	49.5	47.6	ModelName		wCoT(%)	w/oCoT(%)
Qwen2.5-32B-Instruct	50.8	49.9
			o1-mini@20240912		64.0	64.0
Gemma-2-27b-it	36.0	38.9
			GPT-4o-20241120		53.6	51.5
Phi-4-14B	55.3	47.3	LLAMA3.3-70B-Instruct		49.5	47.4
Phi-3-14B	29.5	24.4
			Qwen2.5-32B-Instruct		50.8	44.6
Qwen3-8B	47.5	45.9
			Qwen2.5-7B-Instruct		43.6	42.6
InternLM3-8B-Instruct	41.1	45.8
Qwen2.5-7B-Instruct	43.6	44.5
GLM-4-9b-chat 25.6 32.4 focused modelssuch as o1-mini. We conjecture thatthis
Llama-3.1-8B-Instruct	26.1	23.6
---------------------	----	----	---	---	---	---
discrepancyarisesbecausereasoningmodelsinherentlyuti-
Llama-3.2-3B-Instruct	24.2	24.0
---------------------	----	----	---	---	---	---
lizeCoT-likemechanisms,leadingtotheexplicitaddition
ofCoTredundantandineffective.
3.3.Observationsandfindings
			Consistency	between	English and	Chinese questions.
---	---	---	-----------	-------	-----------	------------------
Multimodalreasoningremainschallengingforcurrent AsshowninFig.5,wetestedtheconsistencyofthesame
questionacrossdifferentlanguagesonR-Bench-T.Itcanbe
models. Wecomparedtheperformanceofthesamemodel
onR-Bench-TandR-Bench-M.Forexample,o1achieved
observedthatmostmodels,suchaso1,Doubao1.5pro,and
69.0%onR-Bench-Tbutonly53.2%onR-Bench-M.The GPT-4o,exhibitacertaindegreeofconsistencyacrossdif-
samesituationalsooccursinothermodels,suchasGPT-4o. ferentlanguages.Thissuggeststhatfoundationmodelshave
Itindicatesthatthemodel’scapabilityinlanguagereasoning alreadydemonstratedacertainlevelofintelligence,enabling
significantlysurpassesitsabilityinmultimodalreasoning. themtoperformreasoningonproblemsofthesamediffi-
Therefore,akeyfocusofresearchintherecentfuturewill cultyindifferentlinguisticenvironments. However,these
behowtotransferlinguisticintelligencetothemultimodal modelsarenotperfectandstillrequirefurtherimprovement
domain. inthisaspect. Thisconsistencyreflectstheextenttowhich
			themodeloverfitsdifferentlanguages.			Therefore,wehope
---	---	---	-----------------------------------	---	---	----------------
futuremodelswillfocusmoreonlearninghowtoreason
TheeffectofCoT. InTab.7,wetestedtheeffectofCoT
----------------------------------------------	---	---	---	---	---	---
onfivemodels. Asseeninthetable,mostmodelsbenefit ratherthanmerelyfittingtospecificlanguages.
fromCoT,butithasnoimpactono1-mini.		Theresultsindi-
----------------------------------	---	---------------	---	---	---	---
catethatCoTenhancestheperformanceofchatmodelslike Modelsshowsignificantperformancevariationacross
GPT-4o. However,ithasnonotableimpactonreasoning- disciplines. Fig.6showstheperformanceofGPT-4oin
7

R-Bench leveragedacrossdiversefieldssuchaswriting,coding,edu- cation,healthcare,finance,andmore,servingasasource Consistency of English and Chinese questions with the same difficulty 86.7%

						forprovidingintelligence.			Now,	foundationmodelshave
	80 78.5%
75.2% 73.5% 73.4% 73.3% becomeanessentialpartofourdailyworkandlife.
			72.4% 70.5%	69.3% 69.2%
---	---	---	-----------	-----------	-----------------	-------------------------------------------------	---	---	---	---	---	---	---
					68.3% 68.0% 66.9%
	60					Tobuildahigh-qualityfoundationmodel,webelievethat
ycnetsisnoC fivekeyaspectsareessential: pre-training(Vaswanietal.,
2017;Raffeletal.,2020;Sunetal.,2023;Radfordetal.,
40
2021;2018),supervisedfine-tuning(Ouyangetal.,2022;
						Liu et al.,	2024b;	Pareja	et	al., 2024;	Li	et al.,	2024;
---	---	---	---	---	---	-----------	------	------	---	----------	---	-------	-----
20
						Zhu et al.,	2023;	Taori	et al.,	2023),	preference	optimiza-
---	---	---	---	---	---	-----------	-----	-----	-------	------	----------	---------	---
tion(Rafailovetal.,2024;Schulmanetal.,2017;Lightman
	0 7 5	2 c t 0	0 c t i t	c t uc t -8B-Instruct	B .5-7B-Instruct
---	-----	-------	---------	---------------------	----------------	---	---	---	---	---	---	---	---
4 1 2 1 4 1 2 1 4 0 9 1 s t r u onnet@ 0 6 2 4 1 1 2 -32B-Inst ru -2 7 b - s tr u s tr 4 -1 4 et al., 2023; Pal et al., 2024; Azar et al., 2024),test-time
	20 2 -2 0 2 @2	0 2 B -I n	20 2 a -2 70B - In	B - In	Ph i-
---	--------------	----------	------------------	-------	-----	---	---	---	---	---	---	---	---
	o1- -p ro w	-7 2 s 4o-	m m -	.1 -8 3
a o e v ie n 2 . 5 3 . 5 - G P T - n 2 . 5 G e a- 3 . 3 a - 3 rn L M e n 2 enhancement(Brownetal.,2020;Weietal.,2022;OpenAI,
	D o u b 1 -p r	Q w e au d e	Q w e la m L la m	nt e	Q w
---	--------------	------------	-----------------	----	---	------------------------------------------------	---	-------	-------	-----	----------	-----	----
	o	C l	L	I		2024b;DeepSeek,2024;Dongetal.,2022),andtrustwor-
						thy evaluation		(Chiang	et al.,	2024;	Li et al.,	2023;	Chen
Figure5.Theperformanceofdifferentmodelsonquestionsofthe
						et al., 2021;	Jain	et al.,	2024;	Yu et	al., 2023).	Reliable
---	---	---	---	---	---	-------------	----	-------	-----	-----	-----------	--------	---
samedifficultyinChineseandEnglish.
evaluationplaysacrucialroleinrevealingmodels’weak-
nessesandshortcomings,guidingfurtheroptimizationand
improvement,whichisalsothefocusofthispaper.
Accuracy of Different Departments
70 68.3%68.0% Average: 53.6% 4.2.Evaluationforfoundationmodels
62.5%
		61.0%				Evaluatingtheintelligenceoffoundationmodelsisamul-
---	---	-----	---	---	---	--------------------------------------------------	---	---	---	---	---	---	---
	60	58.4%
55.9%55.8%55.3%54.5%54.2% tifacetedandcomplexchallenge,akintoassessinghuman
53.6%
)%( ycaruccA 51.0% intelligence. Researchers have introduced various evalu-
	50		48.0%
---	---	---	-----	---	---	-----------------	---	-------	-----------	---	----	-------------	---
						ation benchmarks,		broadly	categorized		into	fast-thinking
44.9%
						assessments	(	a.k.a., system-I		evaluation		(Chiang	et al.,
---	---	---	---	----------	---	------------	---	----------------	-----	----------	----------	-------	-------
	40			39.0%38.9%
						2024; Dubois		et al., 2024;	Zheng	et	al., 2023;	Lin	et al.,
34.4%33.3%
2021;2024;Luetal.,2022;Yuetal.,2023;Lietal.,2023),
30.4%
30
whichrequiresthefoundationmodelstomemorizeexten-
siveknowledgeandretrieveefficiently,andslow-thinking
	Biolog y s E . tronic E	. erospac e hemistry Ma th utomation	s ic s on ic s hanical E. Statist ic s emica l E .	puter S. Materials Civil E	. e E . ironme nt conomics
---	------------------------	------------------------------------	--------------------------------------------------	--------------------------	---------------------------	---	---	---	---	---	---	---	---
hysic Ph y lectr ehicl assessments(a.k.a.,system-IIevaluation(Hendrycksetal.,
	P E l ec	A C A	o e e c C h C om		V E n v E
---	--------	-----	----------------	---	---------	---	---	---	---	---	---	---	---
			M i cr M
2021;Yueetal.,2024a;Luetal.,2023;Wangetal.,2024a;
Liuetal.,2024c;Chenetal.,2021;Jimenezetal.,2023),
R-Bench-T
Figure6.GPT-4o		on	across	different	departments,
--------------	---	---	------	---------	------------	---	---	---	---	---	---	---	---
whichemphasizesthecomplexreasoningskillsoffounda-
whichshowslargevariationamongdifferentdisciplines.. R-Benchfocusesonthelatter.
tionmodels.
MMLU(Hendrycksetal.,2021)isthepioneerinreasoning
different areas of the R-Bench-T benchmark. From the evaluation,whichproposesamulti-disciplineunderstanding
test. Afterthat,lotsofmulti-disciplinebenchmarks(Rein
figure,itcanbeobservedthattheperformancevariessig-
nificantlyacrossdifferentdomains,witharangereaching etal.,2023;Wangetal.,2024b;Yueetal.,2024b)atthe
37.9%. This suggests that if we want to improve the rea- undergraduateorgraduatelevelareproposedforreasoning
						assessment.	However,withtherapiddevelopmentofintelli-
---	---	---	---	---	---	-----------	-----------------------------------------	---	---	---	---	---	---
soningabilityofmodels,weshouldtakeacomprehensive
approachratherthanfocusingsolelyonimprovementsina gentmodels(OpenAI,2024b),thesebenchmarksareclose
						tosaturation.	Besides,severalstudies(Glazeretal.,2024;
---	---	---	---	---	---	-------------	----------------------------------------	---	---	---	---	---	---
singlesubject,suchasmathematics.
Gaoetal.,2024;Luetal.,2023;Wangetal.,2024a)assess
reasoningabilitythroughcomplexmathematicalproblems
4.RelatedWork
						such as	mathematical		olympiad	challenges,		which	bring
---	---	---	---	---	---	-------	------------	---	--------	-----------	---	-----	-----
4.1.Foundationmodels challengesandguidancetocurrentfoundationmodels.How-
ever,onlyguidingthemodeltoimproveitsmathematical
WiththeemergenceofChatGPT(OpenAI,2022),founda-
						reasoning	skills	appears	to be	limited.	In this	paper,	our
---	---	---	---	---	---	---------	------	-------	-----	--------	-------	------	---
tionmodels(Ouyangetal.,2022;Touvronetal.,2023;Yang
targetistobuildareliablebenchmarkforLLMandMLLM
etal.,2024;Jiangetal.,2023;Teametal.,2024;Zengetal., reasoning evaluation, which matches the comprehensive-
2022;	Bi et	al., 2024;	Liu et al., 2024a;	Cai	et al., 2024;
-----	-----	----------	------------------	---	-------------	---	---	---	---	---	---	---	---
nessofMMLU(Hendrycksetal.,2021)whileachieving
Anthropic,2024a;Wuetal.,2024)areincreasinglybeing
8

R-Bench thedifficultyofmathematicalolympiadquestions. Z., Wei, X., Weng, Q., Wu, F., Xiong, Y., Xu, C., Xu, R.,Yan,H.,Yan,Y.,Yang,X.,Ye,H.,Ying,H.,Yu,J.,

5.Conclusion						Yu,J.,Zang,Y.,Zhang,C.,Zhang,L.,Zhang,P.,Zhang,
P.,Zhang,R.,Zhang,S.,Zhang,S.,Zhang,W.,Zhang,
Inthispaper,weproposedR-Bench,agraduate-levelmulti- W.,Zhang,X.,Zhang,X.,Zhao,H.,Zhao,Q.,Zhao,X.,
disciplinary, multilingual benchmark for both LLM and Zhou,F.,Zhou,Z.,Zhuo,J.,Zou,Y.,Qiu,X.,Qiao,Y.,
MLLMreasoningevaluation,whichhascoveragesimilar andLin,D. Internlm2technicalreport,2024.
to MMLU	and MMMU	while reaching		the difficulty	of
-------	--------	--------------	---	--------------	---	---	---	---	---	---	---	---
Chen,M.,Tworek,J.,Jun,H.,Yuan,Q.,Pinto,H.P.D.O.,
mathematicalcompetitionssuchasAIME@2024.				Weeval-
----------------------------------------	---	---	---	-------	---	---	---	---	---	---	---	---
Kaplan,J.,Edwards,H.,Burda,Y.,Joseph,N.,Brockman,
uatedmultipleclosed-sourceandopen-sourcemodelssuch
				R-Bench		G., etal.	Evaluatinglargelanguagemodelstrainedon
---------	-----------	------------	---	-------	---	---------	--------------------------------------	---	---	---	---	---
as OpenAI	o1, GPT-4o,	DeepSekk-R1,		etc, on
code. arXivpreprintarXiv:2107.03374,2021.
andobservedboththeprogressandlimitationsofcurrent
modelsinreasoning. Later,wewillmakethedataandcode Chen, Z., Wang, W., Cao, Y., Liu, Y., Gao, Z., Cui, E.,
available,	hoping to provide	guidance	and	insight	for the
----------	-----------------	--------	---	-------	-------	--------	-------	-----	--------	---------	-------------	---
						Zhu, J.,	Ye, S.,	Tian,	H., Liu,	Z., etal.	Expandingper-
developmentoffoundationmodels.
formanceboundariesofopen-sourcemultimodalmodels
						withmodel,data,andtest-timescaling.					arXivpreprint
----------	---	---	---	---	---	-----------------------------------	---	---	---	---	-------------	---
References						arXiv:2412.05271,2024.
AI, D. Deepseek-r1: Incentivizing reasoning capabil- Chiang,W.-L.,Zheng,L.,Sheng,Y.,Angelopoulos,A.N.,
ity in llms via reinforcement learning. arXiv preprint Li, T., Li, D., Zhang, H., Zhu, B., Jordan, M., Gonza-
arXiv:2501.12948,2025. lez, J. E., et al. Chatbot arena: An open platform for
						evaluating	llms	by human		preference.	arXiv	preprint
---	---	---	---	---	---	----------	----	--------	---	-----------	-----	--------
Anthropic. Theclaude3modelfamily: Opus,sonnet,haiku. arXiv:2403.04132,2024.
https://www.anthropic.com/news/claude-3-family,2024a.
						Contributors,	O.	Opencompass:		A universal		evaluation
----------	-----------	---	------	---	-------	-------------	--------------	------------	-------	---------------	---	----------
Anthropic.	Introducing		claude	3.5	sonnet.					https://github.
						platform	for foundation		models.
https://www.anthropic.com/news/claude-3-5-sonnet, com/open-compass/opencompass,2023.
2024b.
						DeepSeek.		Deepseek-r1		release.		https://api-
---	---	---	---	---	---	---------	---	-----------	---	--------	---	------------
Azar,M.G.,Guo,Z.D.,Piot,B.,Munos,R.,Rowland,M., docs.deepseek.com/news/news250120,2024.
Valko,	M., and Calandriello,	D.	A general	theoretical
------	---------------------	---	---------	-----------	---	---------	-------	----	----------	---	-------	------------
						Dong, Q.,	Li, L.,	Dai,	D., Zheng,	C.,	Ma, J.,	Li, R., Xia,
paradigmtounderstandlearningfromhumanpreferences.
InInternationalConferenceonArtificialIntelligenceand H.,Xu,J.,Wu,Z.,Liu,T.,etal. Asurveyonin-context
						learning.	arXivpreprintarXiv:2301.00234,2022.
---	---	---	---	---	---	---------	-----------------------------------	---	---	---	---	---
Statistics,pp.4447–4455.PMLR,2024.
						Duan, H.,	Yang,	J., Qiao,	Y.,	Fang,	X., Chen,	L., Liu,
---------------------------------	---------	---------	--------	------------	---	-----------	--------------------------------------	---------	----------	-----	---------	----------
Bi, X., Chen,	D., Chen,	G., Chen,	S., Dai,	D., Deng,	C.,
						Y., Dong,	X.,	Zang,	Y., Zhang,		P., Wang,	J., et al.
Ding,H.,Dong,K.,Du,Q.,Fu,Z.,etal.				Deepseekllm:
						Vlmevalkit:	Anopen-sourcetoolkitforevaluatinglarge
Scalingopen-sourcelanguagemodelswithlongtermism.
						multi-modality		models.	In	Proceedings		of the 32nd
---	---	---	---	---	---	--------------	---	-------	---	-----------	---	-----------
arXivpreprintarXiv:2401.02954,2024.
ACMinternationalconferenceonmultimedia,pp.11198–
Brown,T.,Mann,B.,Ryder,N.,Subbiah,M.,Kaplan,J.D.,						11201,2024.
-------------------------------------------------	----------------------------------	---	------	-----------	---	-----------	-------	---	-------	-----------	---	---------
Dhariwal,	P., Neelakantan,	A.,	Shyam,	P., Sastry,	G.,
						Dubey, A.,	Jauhri,	A.,	Pandey,	A., Kadian,	A.,	Al-Dahle,
Askell,A.,etal.	Languagemodelsarefew-shotlearners.
A.,Letman,A.,Mathur,A.,Schelten,A.,Yang,A.,Fan,
Advancesinneuralinformationprocessingsystems,33:
						A., et al.	The	llama	3 herd	of models.	arXiv	preprint
---	---	---	---	---	---	----------	---	-----	------	----------	-----	--------
1877–1901,2020.
arXiv:2407.21783,2024.
Cai,Z.,Cao,M.,Chen,H.,Chen,K.,Chen,K.,Chen,X.,
Dubois,Y.,Galambosi,B.,Liang,P.,andHashimoto,T.B.
Chen,X.,Chen,Z.,Chen,Z.,Chu,P.,Dong,X.,Duan,
						Length-controlled		alpacaeval:		A simple	way	to debias
---	---	---	---	---	---	-----------------	---	-----------	---	--------	---	---------
H.,Fan,Q.,Fei,Z.,Gao,Y.,Ge,J.,Gu,C.,Gu,Y.,Gui,
						automaticevaluators.			arXivpreprintarXiv:2404.04475,
---	---	---	---	---	---	--------------------	---	---	------------------------------	---	---	---
T.,Guo,A.,Guo,Q.,He,C.,Hu,Y.,Huang,T.,Jiang,T.,
2024.
Jiao,P.,Jin,Z.,Lei,Z.,Li,J.,Li,J.,Li,L.,Li,S.,Li,W.,
Li,Y.,Liu,H.,Liu,J.,Hong,J.,Liu,K.,Liu,K.,Liu,X., Gao, B., Song, F., Yang, Z., Cai, Z., Miao, Y., Dong, Q.,
Lv,C.,Lv,H.,Lv,K.,Ma,L.,Ma,R.,Ma,Z.,Ning,W., Li, L., Ma, C., Chen, L., Xu, R., et al. Omni-math:
Ouyang,L.,Qiu,J.,Qu,Y.,Shang,F.,Shao,Y.,Song,D., A universal olympiad level mathematic benchmark for
Song,Z.,Sui,Z.,Sun,P.,Sun,Y.,Tang,H.,Wang,B., largelanguagemodels. arXivpreprintarXiv:2410.07985,
Wang,G.,Wang,J.,Wang,J.,Wang,R.,Wang,Y.,Wang,						2024.
---------------------------------------------	---	---	---	---	---	-----	---	---	---	---	---	---
9

R-Bench Glazer,E.,Erdil,E.,Besiroglu,T.,Chicharro,D.,Chen,E., Lin, S., Hilton, J., and Evans, O. Truthfulqa: Measuring Gunning,A.,Olsson,C.F.,Denain,J.-S.,Ho,A.,Santos, how models mimic human falsehoods. arXiv preprint E.d.O.,etal. Frontiermath: Abenchmarkforevaluating arXiv:2109.07958,2021.

advancedmathematicalreasoninginai.			arXivpreprint
arXiv:2411.04872,2024. Liu,A.,Feng,B.,Wang,B.,Wang,B.,Liu,B.,Zhao,C.,
Dengr,C.,Ruan,C.,Dai,D.,Guo,D.,etal.Deepseek-v2:
Hendrycks,D.,Burns,C.,Basart,S.,Zou,A.,Mazeika,M., A strong, economical, and efficient mixture-of-experts
Song,D.,andSteinhardt,J. Measuringmassivemultitask language model. arXiv preprint arXiv:2405.04434,
language	understanding.	In International	Conference		2024a.
-----------	----------------	----------------	------------	---	------	---	---	---
on Learning	Representations,	2021.	URL https://
openreview.net/forum?id=d7KBjmI3GmQ. Liu,H.,Li,C.,Wu,Q.,andLee,Y.J. Visualinstructiontun-
					ing. Advancesinneuralinformationprocessingsystems,
---------	-----------------	--------------	---------------	---	--------------------------------------------------	---	---	---
Jain, N.,	Han, K., Gu,	A., Li, W.-D.,	Yan, F., Zhang,	T.,	36,2024b.
Wang,	S., Solar-Lezama,	A., Sen,	K., and Stoica,	I.
Livecodebench: Holistic and contamination free eval- Liu,J.,Liu,H.,Xiao,L.,Wang,Z.,Liu,K.,Gao,S.,Zhang,
uationoflargelanguagemodelsforcode. arXivpreprint W., Zhang, S., and Chen, K. Are your llms capable
					ofstablereasoning?	arXivpreprintarXiv:2412.13147,
---	---	---	---	---	------------------	------------------------------	---	---
arXiv:2403.07974,2024.
2024c.
Jiang, A.	Q., Sablayrolles,	A., Mensch,	A., Bamford,	C.,
---------	-----------------	-----------	------------	---	---	---	---	---
Chaplot,D.S.,Casas,D.d.l.,Bressand,F.,Lengyel,G., Lu, P., Mishra, S., Xia, T., Qiu, L., Chang, K.-W., Zhu,
					S.-C., Tafjord,	O., Clark,	P., and Kalyan,	A. Learn to
---------------------------	---	----------	-------------	---	---------------	----------	---------------	-----------
Lample,G.,Saulnier,L.,etal.		Mistral7b.	arXivpreprint
arXiv:2310.06825,2023. explain: Multimodal reasoning via thought chains for
					sciencequestionanswering.		AdvancesinNeuralInfor-
---	---	---	---	---	-------------------------	---	----------------------	---
Jimenez,C.E.,Yang,J.,Wettig,A.,Yao,S.,Pei,K.,Press,
mationProcessingSystems,35:2507–2521,2022.
O.,andNarasimhan,K.		Swe-bench:	Canlanguagemod-
-------------------	---	----------	---------------	---	---	---	---	---
els resolve real-world github issues? arXiv preprint Lu, P., Bansal, H., Xia, T., Liu, J., Li, C., Hajishirzi,
arXiv:2310.06770,2023. H., Cheng, H., Chang, K.-W., Galley, M., and Gao,
					J. Mathvista:	Evaluating	mathematical	reasoning of
---	---	---	---	---	-------------	----------	------------	------------
Kahneman,D. Thinking,fastandslow. Farrar,Strausand foundation models in visual contexts. arXiv preprint
Giroux,2011.					arXiv:2310.02255,2023.
------------	---------------	----------	----------	-------	----------------------	---	-----------	--------
Kwon, W.,	Li, Z., Zhuang,	S., Sheng,	Y., Zheng,	L., Yu,
					OpenAI.		Introducing	chatgpt.
C.H.,Gonzalez,J.E.,Zhang,H.,andStoica,I. Efficient https://openai.com/index/chatgpt/,2022.
memorymanagementforlargelanguagemodelserving
withpagedattention. InProceedingsoftheACMSIGOPS OpenAI. Gpt-4osystemcard. https://openai.com/index/gpt-
4o-system-card/,2024a.
29thSymposiumonOperatingSystemsPrinciples,2023.
Li, B., Wang, R., Wang, G., Ge, Y., Ge, Y., and Shan, Y. OpenAI. Learning to reason with llms.
Seed-bench: Benchmarkingmultimodalllmswithgener- https://openai.com/index/learning-to-reason-with-
llms/,2024b.
ativecomprehension.		arXivpreprintarXiv:2307.16125,
-------------------	---	------------------------------	---	---	---	---	---	---
2023.
Ouyang,L.,Wu,J.,Jiang,X.,Almeida,D.,Wainwright,C.,
Li, B., Zhang, Y., Guo, D., Zhang, R., Li, F., Zhang, H., Mishkin,P.,Zhang,C.,Agarwal,S.,Slama,K.,Ray,A.,
Zhang, K., Zhang, P., Li, Y., Liu, Z., et al. Llava- et al. Training language models to follow instructions
onevision: Easy visual task transfer. arXiv preprint withhumanfeedback. Advancesinneuralinformation
arXiv:2408.03326,2024. processingsystems,35:27730–27744,2022.
Lightman,H.,Kosaraju,V.,Burda,Y.,Edwards,H.,Baker, Pal,A.,Karkhanis,D.,Dooley,S.,Roberts,M.,Naidu,S.,
B., Lee, T., Leike, J., Schulman, J., Sutskever, I., and and White, C. Smaug: Fixing failure modes of pref-
Cobbe, K. Let’s verify step by step. arXiv preprint erence optimisation with dpo-positive. arXiv preprint
arXiv:2305.20050,2023.					arXiv:2402.13228,2024.
----------------------	---	---	---	---	----------------------	---	---	---
Lin,B.Y.,Deng,Y.,Chandu,K.,Brahman,F.,Ravichander, Pareja, A., Nayak, N. S., Wang, H., Killamsetty, K., Su-
A.,Pyatkin,V.,Dziri,N.,Bras,R.L.,andChoi,Y. Wild- dalairaj, S., Zhao, W., Han, S., Bhandwaldar, A., Xu,
bench: Benchmarkingllmswithchallengingtasksfrom G., Xu, K., etal. Unveilingthesecretrecipe: Aguide
realusersinthewild. arXivpreprintarXiv:2406.04770, for supervised fine-tuning small llms. arXiv preprint
2024.					arXiv:2412.13337,2024.
-----	---	---	---	---	----------------------	---	---	---
10

R-Bench Radford,A.,Narasimhan,K.,Salimans,T.,andSutskever, Attentionisallyouneed. InGuyon,I.,Luxburg,U.V., I. Improvinglanguageunderstandingbygenerativepre- Bengio, S., Wallach, H., Fergus, R., Vishwanathan, S., training. 2018. andGarnett,R.(eds.),AdvancesinNeuralInformation ProcessingSystems,volume30.CurranAssociates,Inc., Radford,A.,Kim,J.W.,Hallacy,C.,Ramesh,A.,Goh,G., 2017. URL https://proceedings.neurips. Agarwal,S.,Sastry,G.,Askell,A.,Mishkin,P.,Clark,J., cc/paper_files/paper/2017/file/ etal. Learningtransferablevisualmodelsfromnatural 3f5ee243547dee91fbd053c1c4a845aa-Paper.

language	supervision.		In	International	conference	on	pdf.
machinelearning,pp.8748–8763.PMLR,2021.
Wang,K.,Pan,J.,Shi,W.,Lu,Z.,Zhan,M.,andLi,H.Mea-
Rafailov,R.,Sharma,A.,Mitchell,E.,Manning,C.D.,Er-
							suring multimodal		mathematical	reasoning	with	math-
-------------	---------	-----------	-----------	----------	-------------	---	-----------------	---	------------------------------------	---------	----	-----
mon,	S., and	Finn,	C. Direct	preference	optimization:
							visiondataset.		arXivpreprintarXiv:2402.14804,2024a.
Your language		model	is secretly	a reward	model.	Ad-
vances	in Neural	Information		Processing	Systems,	36,
Wang,Y.,Ma,X.,Zhang,G.,Ni,Y.,Chandra,A.,Guo,S.,
2024.							Ren,W.,Arulraj,A.,He,X.,Jiang,Z.,etal.				Mmlu-pro:
-----	---	---	---	---	---	---	--------------------------------------	---	---	---	---------	---
Amorerobustandchallengingmulti-tasklanguageunder-
Raffel, C.,	Shazeer,	N.,	Roberts,	A., Lee,	K.,	Narang, S.,
-----------------------------------	--------	---	--------	--------	---	-----------	------------------	---	------------------------------	---	---	---
							standingbenchmark.		arXivpreprintarXiv:2406.01574,
Matena,M.,Zhou,Y.,Li,W.,andLiu,P.J.						Exploring
2024b.
thelimitsoftransferlearningwithaunifiedtext-to-text
transformer.		Journalofmachinelearningresearch,				21
------------	---	---------------------------------	---	---	---	---	---	---	---	---	---	---
Wei,J.,Wang,X.,Schuurmans,D.,Bosma,M.,Xia,F.,Chi,
(140):1–67,2020.
E.,Le,Q.V.,Zhou,D.,etal.Chain-of-thoughtprompting
Rein, D., Hou, B. L., Stickland, A. C., Petty, J., Pang, elicitsreasoninginlargelanguagemodels. Advancesin
neuralinformationprocessingsystems,35:24824–24837,
R.Y.,Dirani,J.,Michael,J.,andBowman,S.R.						Gpqa:
----------------------------------------	---	------------	---	---	----------	-----	-----	---	---	---	---	---
A graduate-level		google-proof		q&a	benchmark.	arXiv	2022.
preprintarXiv:2311.12022,2023.
Wu,Z.,Chen,X.,Pan,Z.,Liu,X.,Liu,W.,Dai,D.,Gao,H.,
Schulman, J., Wolski, F., Dhariwal, P., Radford, A., and Ma,Y.,Wu,C.,Wang,B.,etal. Deepseek-vl2: Mixture-
Klimov, O. Proximal policy optimization algorithms. of-experts vision-language models for advanced multi-
arXivpreprintarXiv:1707.06347,2017. modalunderstanding. arXivpreprintarXiv:2412.10302,
2024.
Sun, Q.,	Yu, Q.,	Cui,	Y., Zhang,	F., Zhang,	X.,	Wang, Y.,
--------	-------	----	----------	----------	---	---------	---	---	---	---	---	---
Gao,H.,Liu,J.,Huang,T.,andWang,X. Emu: Genera- Yang, A., Yang, B., Zhang, B., Hui, B., Zheng, B., Yu,
tivepretraininginmultimodality. InTheTwelfthInterna- B.,Li,C.,Liu,D.,Huang,F.,Wei,H.,etal. Qwen2.5
tionalConferenceonLearningRepresentations,2023.
							technicalreport.		arXivpreprintarXiv:2412.15115,2024.
----------	----------	---	----------	-----------	---	-----------	----------------	---	-----------------------------------	---	---	---
Taori, R.,	Gulrajani,		I., Zhang,	T., Dubois,		Y., Li, X.,
Yu,W.,Yang,Z.,Li,L.,Wang,J.,Lin,K.,Liu,Z.,Wang,
Guestrin,C.,Liang,P.,andHashimoto,T.B. Stanford X., and Wang, L. Mm-vet: Evaluating large multi-
alpaca:	Aninstruction-followingllamamodel,2023.
-------	---------------------------------------	---	---	---	---	---	-------------------------------------	---	---	---	-------------	---
							modalmodelsforintegratedcapabilities.				arXivpreprint
arXiv:2308.02490,2023.
Team,G.,Anil,R.,Borgeaud,S.,Alayrac,J.-B.,Yu,J.,Sori-
cut,R.,Schalkwyk,J.,Dai,A.M.,Hauth,A.,Millican, Yue,X.,Ni,Y.,Zhang,K.,Zheng,T.,Liu,R.,Zhang,G.,
K.,etal.	Gemini:	afamilyofhighlycapablemultimodal
--------	-----------------------------------	--------------------------------	---	---	---	---	---------------------------------------	---	---	---	---	-------
							Stevens,S.,Jiang,D.,Ren,W.,Sun,Y.,etal.					Mmmu: A
models.	arXivpreprintarXiv:2312.11805,2023.
massivemulti-disciplinemultimodalunderstandingand
Team, G., Georgiev, P., Lei, V. I., Burnell, R., Bai, L., reasoningbenchmarkforexpertagi.InProceedingsofthe
IEEE/CVFConferenceonComputerVisionandPattern
Gulati,	A., Tanzer,		G., Vincent,	D.,	Pan, Z.,	Wang, S.,
-------	-----------	---	------------	---	--------	---------	---	---	---	---	---	---
Recognition,pp.9556–9567,2024a.
et al.	Gemini	1.5:	Unlocking	multimodal		understand-
------	------	----	---------	----------	---	-----------	---	---	---	---	---	---
arXivpreprint
ingacrossmillionsoftokensofcontext.
Yue,X.,Zheng,T.,Ni,Y.,Wang,Y.,Zhang,K.,Tong,S.,
arXiv:2403.05530,2024.
							Sun,Y.,Yu,B.,Zhang,G.,Sun,H.,etal.				Mmmu-pro:	A
---	---	---	---	---	---	---	----------------------------------	---	---	---	---------	---
Touvron,H.,Lavril,T.,Izacard,G.,Martinet,X.,Lachaux, morerobustmulti-disciplinemultimodalunderstanding
M.-A.,Lacroix,T.,Rozie`re,B.,Goyal,N.,Hambro,E., benchmark. arXivpreprintarXiv:2409.02813,2024b.
Azhar,F.,etal.		Llama:Openandefficientfoundationlan-
--------------	---	------------------------------------	---	---	---	---	-----------------------------------	---	-------------	---	---------	---------
							Zeng, A., Liu,	X.,	Du, Z., Wang,	Z.,	Lai, H.,	Ding, M.,
guagemodels.		arXivpreprintarXiv:2302.13971,2023.
							Yang,Z.,Xu,Y.,Zheng,W.,Xia,X.,etal.				Glm-130b:
Vaswani,A.,Shazeer,N.,Parmar,N.,Uszkoreit,J.,Jones, An open bilingual pre-trained model. arXiv preprint
L., Gomez, A. N., Kaiser, L. u., and Polosukhin, I. arXiv:2210.02414,2022.
11

R-Bench | Zheng,L.,Chiang,W.-L.,Sheng,Y.,Zhuang,S.,Wu,Z., | | | | A.Appendix. | | | | | | | | ----------------------------------------------- | --- | --- | ------- | ----------- | --- | --- | --- | --- | --- | --- | | Zhuang,Y.,Lin,Z.,Li,Z.,Li,D.,Xing,E.,etal. | | | Judging | | | | | | | | A.1.ResponseExample | llm-as-a-judge | with mt-bench | and chatbot | arena. Ad- | | | | | | | | | -------------- | ------------- | ----------- | ---------- | --- | --- | --- | --- | --- | --- | --- | vances in Neural Information Processing Systems, 36: Here, wepresentourquestionandtheoutputoftheOpe- 46595–46623,2023. | | | | | nAI’so1modelintheboxbelow. | | | | Themodelcarriedouta | | | | --- | --- | --- | --- | -------------------------- | --- | --- | --- | ------------------- | --- | --- | step-by-stepanddetailedreasoningbasedonthequestion | Zhu, D., | Chen, J., Shen, | X., Li, X., | and Elhoseiny, M. | | | | | | | | | -------- | --------------- | ----------- | ----------------- | --- | --- | --- | --- | --- | --- | --- | andfinallyarrivedatthecorrectanswer. | Minigpt-4: | Enhancing | vision-language | understanding | | | | | | | | | ------------- | -------------- | --------------- | -------------- | --- | --- | --- | --- | --- | --- | --- | | with advanced | large language | models. | arXiv preprint | | | | | | | | Questionandtheresponsebyo1 arXiv:2304.10592,2023. | | | | | Answer | the | following | single choice | question. | | The | | --- | --- | --- | --- | ------ | --- | --------- | ------------- | --------- | --- | --- | lastlineofyourresponseshouldbeofthefollowing | | | | | format: | ’ANSWER: | | $LETTER’ | (without | quotes) | | | --- | --- | --- | --- | ----------------------------------------- | ----------- | ---------- | --------------- | -------- | ---------- | ------ | | | | | | where | LETTER | is one | of Options(e.g. | | | one of | | | | | | ABCDEF). | | Think step | by step | before | answering. | | | | | | | Question: | | A sudden | environmental | change | | causes | | | | | | arecessivealleletobecomecompletelylethal. | | | | | | How | | | | | | many | generations | will | it take | for its | frequency | to | | | | | | decrease | from | 10−2 to | 10−3? | | | | A. 950 B. 850 | | | | | C. | All other | answers are | incorrect | | | | | --- | --- | --- | --- | --- | --------- | ----------- | --------- | --- | --- | --- | D. 900 E. 800 F. 920 Answer: | | | | | ### | Response | by o1 | | | | | | --- | --- | --- | --- | ---------------------------------- | ----------- | ------- | --------- | -------------- | --------- | --- | | | | | | To | solve this, | use the | fact that | for a fully | recessive | | | | | | | lethalallele(withinitialfrequencyq | | | | ),itsfrequency | | | 0 | | | | | in the | next | generation | is given | by: | | | | --- | --- | --- | --- | ------ | ---- | ---------- | -------- | --- | --- | --- | q | | | | | | | q | = n | | | | | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | n+1 1+q n 1 | | | | | Define | R n | = . Then | from the | recurrence | | above: | | --- | --- | --- | --- | ------ | --- | -------- | -------- | ---------- | --- | ------ | qn 1 | | | | | | | R | = | | | | | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | | | | | | | | n+1 | q | | | | n+1 1+q n

q n 1

							= +1
q n
							=R +1
---	---	---	---	---	---	---	-----	---	---	---
n
				Thus,R	=R	+n. SinceR		= 1 andwewantq
---	---	---	---	-----------	-------	-----------	---------	--------------	---	---
					n	0	0	q0		n
				to decrease		from 0.01	to 0.001:
				•	R =	1 =100.
					0	0.01
				•	We need	1 =1000,	so R	=1000.
						qn		n
				•	Hence	1000=100+n,	giving	n=900.
				ANSWER:		D
A.2.CoTPrompt
Here,wepresenttheformatofourCoTpromptinfollowing
box. Aswecansee,ourCoTpromptmainlyuses“Think
stepbystep”,whichshowsthatevenasimplepromptstill
12

R-Bench hasapositiveeffectonmostchatmodels. SystemPromptforR-Bench Answer the following single-choice question. The lastlineofyourresponseshouldbeofthefollowing format: ’ANSWER: $LETTER’ (without quotes) where LETTER is one of the Options (e.g. one of ABCDEF). Think step by step before answering. Question: {Question Input} A. {OptionA} B. {OptionB} C. {OptionC} D. {OptionD} E. {OptionE} F. {OptionF} Answer:

Example:

Answer the following single choice question. The lastlineofyourresponseshouldbeofthefollowing format: ‘ANSWER: $LETTER’ (without quotes) where LETTER is one of Options(e.g. one of ABCDEF). Think step by step before answering. Question: ConsideraCMOSinverterdrivingawire of length L. In the initial design, the on-resistance oftheinverterisequaltothetotalresistanceofthe wire, the source-drain capacitance of the inverter is equal to the total capacitance of the wire, and the total delay of the inverter and wire is tp. Now, the devices are scaled down using Constant Field Scaling, while the wire is ideally scaled down. Assumingthewirecanbemodeledusingalumped parameter model, answer the following questions in a first-order approximation: (1) Assuming the wire is a local wire, and the scaling factors for both process and supply voltage are 2, express the total delay after scaling in terms of tp. (2) Now assume the wire is global, and the length of the wire increases inversely with the process scaling, with scaling factors for both process and supply voltage being 2, express the total delay after scaling in terms of tp. A. (1) 1/3tp (2) 35/6tp B. (1) 1/2tp (2) 38/6tp C. All other answers are incorrect D. (1) 3/4tp (2) 36/6tp E. (1) 5/6tp (2) 39/6tp F. (1) 2/3tp (2) 37/6tp Answer: A.3.Specificsubjectdistribution WepresentthespecificsubjectdistributionsofR-Bench-T andR-Bench-MinTable8andTable8,respectively. Itcan beobservedthatR-Benchhasabroadcoverage,makingit difficulttoimproveperformanceonR-Benchbyoverfitting tospecificsubjects. 13

R-Bench DistributionofCoursesbyDisciplineinR-Bench-T Table8:

Discipline	SpecificSubject	Count
	FluidMechanics	33
CivilEngineering	StructuralMechanics	2
	Surveying	1
	PrinciplesofProcessTransport	3
ChemicalEngineering	PrinciplesofChemicalEngineering	15
	AnalyticalChemistry	7
	AdvancedMathematicalEconomics	1
	Econometrics	3
	IntermediateFinancialTheory	1
Economics	FinancialEngineering	1
	GameTheoryandMechanismDesign	5
	IntermediateMicroeconomics	2
	PrinciplesofAccounting	2
	TimeSeriesAnalysis	8
	SoilScience	1
	Genetics	16
Biology
	Physiology	1
-------	-----------------------------------	---
	Biochemistry	15
	Heredity	8
	MathematicalMethodsinPhysics	10
	Electromagnetics	11
	Optics	23
	QuantumMechanics	8
Physics	AnalyticalMechanics	6
	Electrodynamics	5
	ThermodynamicsandStatisticalPhysics	6
	GeneralRelativity	2
	BasicPhysics	29
	GroupTheory	3
	MechanicsofMaterials	1
	StructuralMechanicsofAircraft	3
	TheoreticalMechanics	2
Aerospace
	FluidMechanicsandAerodynamics	15
---	-----------------------------------------	---
	OptimalControl	29
	PropulsionPrinciplesandThermalFluidBasics	9
	DigitalLarge-ScaleIntegratedCircuits	20
Microelectronics
	AnalogCircuits	2
----------	-------------------------------	---
	SignalsandSystems	20
Automation	OperationsResearch	15
	AutomaticControlTheory	17
	CommunicationandNetwork	15
	PrinciplesofAnalogCircuits	4
	ElectromagneticFieldsandWaves	8
	FundamentalsofSolidStatePhysics	12
ElectronicEngineering
	DigitalSignalProcessing	8
---	-----------------------	---
	StochasticProcesses	3
Continuedonnextpage
14

R-Bench Table8–Continuedfrompreviouspage

Discipline	SpecificSubject	Count
	SolidStatePhysics	4
	AppliedStochasticProcesses	26
	FluidMechanics	11
	PrinciplesandInterfaceTechnologyofSingle-ChipMicrocomputer	4
	MechanicalDesign	1
	TheoryofMachines	4
	ElectromechanicalTransmissionandControl	4
MechanicalEngineering
	ElectricalandelectronicTechnology	2
---------------	------------------------------------	---
	MechanicalVibration	3
	HydraulicandPneumaticTransmission	1
	EngineeringThermodynamics	8
	MechanicsofMaterials	9
	TheoreticalMechanics	1
	DataStructure	24
	CombinatorialMathematics	3
	NumericalAnalysis	5
	Cryptography	17
ComputerScience	Automata	8
	PrinciplesofComputerOrganization	2
	CompilationPrinciples	5
	ComputerNetwork	11
	OperatingSystem	11
	ComputerArchitecture	3
	InorganicChemistry	50
	ChemicalThermodynamics	48
	ChemicalReactionKinetics	20
	IntroductiontoComputationalChemistry	11
Chemistry
	QuantumChemistry	5
---	----------------------------	---
	PhysicalChemistry	22
	OrganicChemistry	5
	ComplexAnalysis	26
	AnalyticGeometry	26
	AdvancedCalculus	9
	NumberTheory	22
	MatrixAnalysis	36
	PartialDifferentialEquations	29
Mathematics
	MathematicalAnalysis	9
---	-----------------------------------	---
	StochasticDifferentialEquations	8
	FunctionalAnalysis	14
	OrdinaryDifferentialEquations	3
	DifferentialGeometry	3
	Topology	3
	ThermodynamicsandStatisticalPhysics	5
PhysicsEngineering
	NuclearRadiationPhysicsandDetection	20
---	-----------------------------------	---
	QuantumandStatistics	14
	PhysicalPropertiesofMaterials	1
Materials
	FundamentalsofMaterialsScience	23
---	------------------------------	---
Continuedonnextpage
15

R-Bench Table8–Continuedfrompreviouspage

Discipline	SpecificSubject	Count
	MaterialsAnalysisandCharacterization	3
	FiniteElementAnalysisBasics	1
	PrinciplesofAutomotivePowerSystem	11
	AutomotiveElectronicsandControl	1
VehicleEngineering	FundamentalsofControlEngineering	3
	TheoryofAutomobile	3
	AutomobileConstruction	1
	DiscreteMathematics	12
	ProbabilityTheory	42
	IntroductiontoBayesianStatistics	6
Statistics
	ReliabilityDataandSurvivalAnalysis	1
-----------	---------------------------------------------------------	---
	StatisticalInference	2
	PrinciplesofEnvironmentalEngineeringScienceandEngineering	5
Environment	WaterTreatmentEngineering	12
	EnvironmentalChemistry	1
16

R-Bench DistributionofCoursesbyDisciplineinR-Bench-M Table9:

Discipline	SpecificSubject	Count
	MaterialsMechanics	26
	FluidMechanicsandAerodynamics	2
	TheoreticalMechanics	16
Aerospace
	AircraftStructuralMechanics	2
---	------------------------------------	---
	PropulsionPrinciplesandThermalFluids	2
	OptimalControl	2
	DigitalVLSI	11
	AnalogCircuits	19
IntegratedCircuits
	DigitalElectronicsFundamentals	8
-------------------	------------------------------	---
	AnalogElectronicsFundamentals	6
	TransportProcessPrinciples	2
ChemicalEngineering	ChemicalPrinciples	11
	PhysicalChemistry	12
	ChemicalThermodynamics	1
	ChemicalReactionKinetics	10
Chemistry	InorganicChemistry	1
	OrganicChemistry	17
	PhysicalChemistry	2
	DataStructures	4
	Combinatorics	1
	DiscreteMathematics	12
	TheoryofAutomata	7
ComputerScience	OperatingSystems	6
	Compilers	4
	ComputerArchitecture	3
	Cryptography	1
	ComputerNetworks	1
	AnalyticalMechanics	25
	Optics	6
Physics	Electrodynamics	9
	Electromagnetism	10
	BasicPhysics	10
	AnalogCircuitPrinciples	17
	SignalsandSystems	3
	DigitalSignalProcessing	1
ElectricalEngineering
	CommunicationandNetworks	4
-----------	-----------------------------	---
	ElectromagneticFieldsandWaves	1
	SolidStatePhysics	1
	ComplexAnalysis	8
	AnalyticGeometry	2
	ProbabilityTheory	5
	StochasticProcesses	3
Mathematics	Analysis	1
	ProbabilityandStatistics	1
	MathematicalAnalysis	5
	Statistics	6
Continuedonnextpage
17

R-Bench Table9–Continuedfrompreviouspage

Discipline	SpecificSubject	Count
	Topology	1
	WaterTreatmentEngineering	4
	EnvironmentalScienceandEngineeringPrinciples	9
	EnvironmentalMonitoring	6
EnvironmentalEngineering
	WaterPollutionControlProject	4
---	------------------------------	---
	EnvironmentalChemistry	1
	SolidWasteTreatmentandDisposal	2
	Genetics	13
Biology
	Biochemistry	4
---	------------------------------	---
	MaterialsMechanics	37
	QuantumandStatisticalMechanics	5
MaterialScience
	MaterialAnalysisandCharacterization	1
---------	---------------------------------------	---
	BasicMaterialScience	18
	OperationsResearch	2
	PrinciplesofAccounting	3
Economics	FinancialEngineering	2
	IntermediateFinancialTheories	3
	GameTheoryandMechanismDesign	8
	ElectricalandElectronicsTechnology	5
	MechanicalDesign	24
	ElectromechanicalTransmissionandControl	1
	HydraulicandPneumaticTransmission	9
MechanicalEngineering
	MechanicalVibrations	7
------------------	-------------------------------	---
	FluidMechanics	7
	PrinciplesofMechanics	1
	TheoreticalMechanics	11
CivilEngineering	StructuralMechanics	33
EngineeringPhysics	EngineeringMechanics	23
Automation	AutomaticControlTheory	25
	FluidMechanics	34
	EngineeringControlBasics	8
	FiniteElementAnalysisBasics	5
VehicleEngineering	AutomotiveElectronicsandControl	5
	AutomobileConstruction	5
	AdvancedHeatTransfer	9
	AutomotivePowerSystemPrinciples	2
Architecture	StructuralEngineering	21
18