Datasets:

luoojason
/

muddle-eval-bundle-md

	DUET: Joint Exploration of User–Item Profiles in Recommendation System
	YueChen1,,YifeiSun1,,LuWang2,†,FangkaiYang2,PuZhao2,MinjieHong3,YifeiDong4,MinghuaHe2,
	NanHu2,JianjinZhang2,ZhiweiDai2,YuefengZhan2,WeihaoHan2,HaoSun2,
	QingweiLin2,WeiweiDeng2,FengSun2,QiZhang2,SaravanRajmohan2,DongmeiZhang2
	1PekingUniversity2Microsoft3ZhejiangUniversity4KTHRoyalInstituteofTechnology
	*Equalcontribution†Correspondingauthor
	Abstract
	\| \| \| \| User History: \| Item History: \| User History: \| Item History: \|
	\| --- \| --- \| --- \| -------------------- \| ------------------------- \| -------------------- \| ------------------------- \|
	\| \| \| \| • Misfits (Punk) \| • Review: \| • Misfits (Punk) \| • Review: \|
	\| \| \| \| • I r o n M a iden \| " P o l i s he d P o p " \| • I r o n M a iden \| " P o l i s he d P o p " \|
	Traditionalrecommendationsystemsrepresent ( M e t al ) • R e v i e w : " G r e a t ( M e t al ) • R e v i e w : " G r e a t
	6202 rpA 51 ]RI.sc[ 1v10831.4062:viXra • • Parliament Funk Bass"
	\| usersanditemsasdensevectorsandlearnto \| \| \| Parliament \| Funk Bass" \| \| \|
	\| ------------------------------------- \| --- \| --- \| ----------- \| ------------------ \| ------ \| ------------------ \|
	\| \| \| \| (Funk) \| • Meta: "80s Rock" \| (Funk) \| • Meta: "80s Rock" \|
	aligntheminasharedlatentspaceforrelevance
	\| \| \| \| \| Duet \| Independent \| \|
	\| --- \| --- \| --- \| --- \| ----- \| ------------ \| --- \|
	estimation. RecentLLM-basedrecommenders (Joint Alignment) Generation
	\| instead leverage \| natural-language \| representa- \| \| \| \| \|
	\| ---------------- \| ---------------- \| ----------- \| ------------- \| ------------- \| ------------- \| ------------- \|
	\| \| \| \| User Profile: \| Item History: \| User Profile: \| Item History: \|
	tionsthatareeasiertointerpretandintegrate Enthusiast of A peak-era FUNK- Fan of HEAVY A highly rated POP-
	\| \| \| \| FUNK and SOUL. \| ROCK album with \| METAL and \| ROCK album with \|
	\| --- \| --- \| --- \| --------------- \| ---------------- \| ---------- \| ---------------- \|
	withdownstreamreasoningmodules. Thispa- Values technical genre-blurring PUNK. Prefers polished
	\| \| \| \| mastery \| creativity. \| dark themes. \| production. \|
	\| --- \| --- \| --- \| ------- \| ----------- \| ------------ \| ----------- \|
	textual
	per studies how to construct effective Semantic Alignment ✘ Semantic Mismatch
	profilesforusersanditems,andhowtoalign
	themforrecommendation. Acentraldifficulty Figure1: DUETalignsrawuseranditemdatabytrans-
	\| \| \| \| forming them \| into textual \| profiles within \| a shared se- \|
	\| --- \| --- \| --- \| ------------ \| ------------ \| --------------- \| ------------ \|
	isthatthebestprofileformatisnotknownapri-
	manticspace.
	ori: manuallydesignedtemplatescanbebrittle
	\| andmisalignedwithtaskobjectives. \| \| Moreover, \| \| \| \| \|
	\| -------------------------------- \| --- \| --------- \| ------------ \| ------------ \| -------------------- \| --- \|
	\| \| \| \| to introduce \| semantically \| rich, human-readable \| \|
	generatinguseranditemprofilesindependently
	mayproducedescriptionsthatareindividually representationsforrecommendation(Wangetal.,
	plausible yet semantically inconsistent for a 2025;Zhang,2024;Baoetal.,2023;Hongetal.,
	\| specificuser–itempair. \| WeproposeDUET,an \| \| \| \| \| \|
	\| ---------------------- \| ---------------- \| --- \| -------------------------- \| --- \| ------------------- \| --- \|
	\| \| \| \| 2025a,b,c;Wangetal.,2024). \| \| Anaturaldirectionis \| \|
	interaction-awareprofilegeneratorthatjointly
	toreplacelatentvectorswithtextualuseranditem
	producesuseranditemprofilesconditionedon
	profilesthatcanbeinspected,edited,andreusedby
	\| both user history \| and item evidence. \| DUET \| \| \| \| \|
	\| ----------------- \| ------------------ \| ---- \| --- \| --- \| --- \| --- \|
	downstreamcomponentsincorporatedwithLLMs.
	\| follows a three-stage \| procedure: \| it first turns \| \| \| \| \|
	\| --------------------- \| ---------- \| -------------- \| --- \| --- \| --- \| --- \|
	However,existingLLM-basedapproachesremain
	rawhistoriesandmetadataintocompactcues,
	then expands these cues into paired profile limited in two important ways. First, directly
	promptsandthengenerateprofiles,andfinally promptinganLLMwithrawuseranditemhisto-
	optimizesthegenerationpolicywithreinforce-
	riestoobtainrecommendationsoftenyieldsnoisy
	mentlearningusingdownstreamrecommenda- andincompletesignals,especiallywhenhistories
	\| tionperformanceasfeedback. \| \| Experimentson \| \| \| \| \|
	\| -------------------------- \| --- \| ------------- \| --------- \| ------------------------ \| --- \| ------------- \|
	\| \| \| \| are long, \| sparse, or heterogeneous \| \| (Wang et al., \|
	threereal-worlddatasetsshowthatDUETcon-
	\| \| \| \| 2025). Second, \| profile-based \| methods \| typically \|
	\| --- \| --- \| --- \| -------------- \| ------------- \| ------- \| --------- \|
	sistentlyoutperformsstrongbaselines,demon-
	relyonmanuallydesignedtemplatesorhandcrafted
	stratingthebenefitsoftemplate-freeprofileex-
	attributes,whichrequiressubstantialhumanengi-
	plorationandjointuser–itemtextualalignment.
	Project code page: https://github.com/duet- neering and constrains the representation space.
	review/duet_code. More fundamentally, many approaches generate
	userprofilesanditemprofilesindependently,with-
	1 Introduction
	\| \| \| \| out modeling \| how user \| preferences \| and item se- \|
	\| --- \| --- \| --- \| ------------ \| -------- \| ----------- \| ------------ \|
	Traditional recommendation systems represent manticsinteractatdecisiontime(Yangetal.,2023;
	Xietal.,2024).
	usersanditemsasdensevectorsandlearntoalign
	theminasharedlatentspaceforrelevanceestima- Toaddressthesechallenges,wepropose DUET,
	tion (Covington et al., 2016; Wu, 2023). While ajointuser–itemprofilegeneratorthattakesboth
	effective,suchembeddingsareopaque: theyoffer userhistoryanditemhistoryasinputandproduces
	limitedinterpretabilityandmakeitdifficulttoan- a paired set of profiles for the interaction. Cru-
	alyzewhyanitemisrecommended. Recentwork cially, DUET doesnotrequireprofiletemplates: it
	thereforeleverageslargelanguagemodels(LLMs) istrainedwithreinforcementlearningusingfeed-

	back from downstream recommendation perfor- asseeninworkslikeCRESDUP(Chenetal.,2007)
	mance,enablingittoexploreanddiscovereffective andUPCSim(Widiyaningtyasetal.,2021). While
	profileformatsautomatically. Figure1illustrates foundational,thesemethodswerelimitedbytheir
	whyjointprofilingmatters. Theuserhaslistened rigid,hand-engineeredfeatures. Morerecently,the
	toMisfits(punk),IronMaiden(metal),andParlia- adventofLargeLanguageModels(LLMs)hasen-
	ment(funk),whilethecandidatealbumisdescribed abledashifttowardsgeneratingprofilesinnatural
	byreviewssuchas“polishedpop”and“greatfunk language. Studies such as KAR (Xi et al., 2024),
	bass”andametatag“’80srock.” Consideringboth GPG(Zhang,2024),PALR(Yangetal.,2023),and
	sides together, DUET can reconcile these signals LettinGo (Wang et al., 2025) leverage LLMs to
	intoacompatibleinterpretation(e.g.,highlighting createtextualuserprofilesfrombehavioraldata. A
	the user’s funk/soul affinity and the item’s funk- keylimitationoftheseapproachesistwofold: they
	rockcharacter). Incontrast,independentlygener- relyonstaticorpre-definedtemplatesforprofile
	atedprofilesmayamplifydifferentfacets,summa- generation,andtheyoftenfocusexclusivelyonuser
	rizingtheuseras“heavymetal/punk”whilesum- profiles,neglectingtherich,expressiveinformation
	marizingtheitemas“pop-rock”,whichresultingin inherentinitemsandthecomplexdynamicsofuser-
	asemanticallymismatchedpairthatobscuresthe iteminteractions. Incontrast,ourworkintroduces
	truerelevancesignal. anewparadigmwherebothuseranditemprofiles
	DUETproceedsinthreestages. First,rawhisto- arenotfixedbutaredynamicallyexploredandop-
	ries and metadata are distilled into minimal cues timizedinasharedsemanticspacetodirectlyalign
	thatcapturecompactbutinformativesignals. Sec- withrecommendationperformance.
	ond, the model expands cues into richer profile
	prompts to generate textual profiles, allowing ex- 2.2 ReinforcementLearningforLLM-Based
	plorationoveralternativeprofilestructuresandem- RecommendationSystems
	phases. Third,theresultingprofilesareconsumed
	ReinforcementLearning(RL),particularlythrough
	bydownstreamrecommenders,andtheirtaskfeed-
	techniqueslikeRLHF,hasbecomeacoremethod
	backisusedtooptimizetheprofilegenerationpol-
	for aligning LLMs with specific objectives. This
	icy. By coupling both sides in a shared semantic
	approach has been adapted for recommendation
	space,DUETlearnsuserprofilesthatreflectwhat
	tasks (Wang et al., 2025; Lin et al., 2025; Deng
	kindsofitemsauserprefersanditemprofilesthat
	etal.,2025),butexistingeffortsoftenfacekeylimi-
	reflectwhatkindsofusersanitemappealsto.
	tations. Theyfrequentlyrelyonofflinerewardmod-
	Ourcontributionsareasfollows:
	els(Jeongetal.,2023;Chenetal.,2025;Heetal.,
	• Werepresentusersanditemsasnatural-language 2025; Liu et al., 2025) that do not adapt in real-
	profiles and align them in a shared semantic timetosystemfeedback,asetupthatrisksissues
	space,extendingtheclassicvector-basedalign- like reward hacking (Skalse et al., 2025). Other
	mentprincipletointerpretabletextualrepresenta- methods(Sunetal.,2024;Luetal.,2024)restrict
	tions. themselvestoofflinepreferencetuning(e.g.,DPO),
	• We introduce an exploration-based framework which can easily overfit on static datasets. Our
	thatstartsfromcue-basedinitialization,expands framework, DUET,overcomesthesechallengesby
	cuesintocandidateprofiles,andjointlyoptimizes integrating RL into a closed-loop system where
	user and item profiles with downstream recom- downstreamrecommendationperformanceserves
	mendationfeedbackviareinforcementlearning, asthereal-timerewardsignal,allowingforthedy-
	avoidingrigidtemplates. namicandinteractiverefinementoftextualprofiles.
	• Extensiveexperimentsacrossmultiplereal-world
	datasets show that DUET consistently outper- 3 Method
	formsstrongbaselines,validatingbothjointpro-
	filingandfeedback-drivenprofileoptimization. DUET is a closed-loop framework that trans-
	forms raw user–item interaction histories into
	2 RelatedWork performance-aligned textual profiles through
	learnedrepresentationstrategies. AsshowninFig-
	2.1 ProfilesinRecommendation
	ure 2, DUET consists of three stages. (1) Cue-
	Earlyrecommendationsystemsprimarilyreliedon Based Initialization distills interaction histories
	pre-definedprofilesbasedonstructuredattributes, intoconciseevidence-basedcues. (2)JointExplo-

	CUE-BASED INITIALIZATION JOINT EXPLORATION ON-POLICY OPTIMIZATION
	\| \| \| \| \| Item \| DU \| \| \| Duet \| \| \| \|
	\| ----------- \| --- \| --- \| --- \| ---- \| --- \| --- \| --- \| ---- \| --- \| --- \| --- \|
	\| ItemHistory \| \| \| \| \| ET \| \| \| \| \| \| \|
	Initial CUE
	Rated: Baba Is You (5★)
	User: “Prefers retro puzzle games”
	Reviewed: “I love logic-
	Item: “Retro-style indie puzzle game”
	\| \| \| \| based puzzle games.” \| \| \| Profile Prompt \| \| \| \| \| \|
	\| --- \| --- \| --- \| -------------------- \| --- \| --- \| -------------- \| --- \| --- \| --- \| --- \| --- \|
	Rated: The Witness (5★)
	User: “Describe the user's typical
	\| \| \| \| Rated: Fez (4★) \| \| puzzle-game preferences … \| \| \| 𝑃𝑟𝑜𝑓𝑖𝑙𝑒1 𝑃𝑟𝑜𝑓𝑖𝑙𝑒2 \| … \| 𝑃𝑟𝑜𝑓𝑖𝑙𝑒𝑛 \| \|
	\| --- \| --- \| --- \| --------------- \| --- \| ------------------------- \| --- \| --- \| ----------------- \| --- \| -------- \| --- \|
	Item: “… item's core mechanics…
	\| \| \| \| \| \| \| \| Profile \| Downstream Task Env \| \| \| \|
	\| ----------- \| --- \| --- \| ----------------------- \| ---- \| --------------------------------- \| --- \| ------- \| --------------------- \| --- \| --- \| --- \|
	\| \| \| \| \| User \| User: “A player who enjoys logic- \| \| \| \| \| \| \|
	\| UserHistory \| \| \| \| \| intensive indie puzzle games, \| \| \| \| \| \| \|
	\| \| \| \| Category: Indie Puzzle \| \| \| \| \| Ranking correctness \| \| \| \|
	prefers retro visual aesthetics…”
	\| \| \| \| Game \| \| \| \| \| &format Reward \| \| \| \|
	\| --- \| --- \| --- \| ---- \| --- \| --- \| --- \| --- \| -------------- \| --- \| --- \| --- \|
	Reviews:
	Item: “A retro-style indie puzzle
	\| \| \| \| “Challenging logic \| \| game featuring challenging \| \| \| Group Computation \| \| \| \|
	\| --- \| --- \| --- \| -------------------- \| ------------------------ \| ------------------------------- \| --- \| --- \| ------------------- \| --- \| ---- \| --- \|
	\| \| \| \| mechanics.” \| \| logic mechanics that appeal to \| \| \| \| \| \| \|
	\| \| \| \| “ R e t r o p ix \| e l- a rt st y l e . ” \| hardcore puzzle players.” \| \| \| 𝐴𝑑𝑣1 𝐴𝑑𝑣2 \| \| \| \|
	\| \| \| \| “ H a r d b u \| t re w a r d i n g . ” \| \| \| \| \| … \| 𝐴𝑑𝑣𝑛 \| \|
	\| \| \| \| \| Figure2: \| OverviewoftheDUETframework. \| \| \| \| \| \| \|
	ration via Adaptive Profile Prompt Discovery 3.2 Cue-BasedInitialization
	\| jointly \| explores \| user-item’s \| \| profile prompts \| that \| \| \| \| \| \| \|
	\| ---------- \| -------- \| ----------- \| -------- \| --------------- \| -------- \| -------- \| --------- \| -------- \| -------- \| --- \| --------- \|
	\| \| \| \| \| \| \| Raw user \| histories \| and item \| metadata \| \| are often \|
	\| define how \| user \| and item \| profiles \| should \| be writ- \| \| \| \| \| \| \|
	noisy,redundant,andnotdirectlysuitableforpro-
	ten. (3)OptimizationviaOn-policyExploration
	\| \| \| \| \| \| \| fileconstruction. \| \| Toaddressthis,DUETintroduces \| \| \| \|
	\| --- \| --- \| --- \| --- \| --- \| --- \| ----------------- \| --- \| ---------------------------- \| --- \| --- \| --- \|
	jointlyoptimizesuseranditemprofilesunderdown-
	theconceptofcues:
	concisehypothesesthatsum-
	streamrecommendationfeedback.
	\| \| \| \| \| \| \| marize minimal \| \| but informative \| \| aspects \| of users \|
	\| ------------------- \| ------ \| --- \| ------------------------- \| ------- \| -------- \| -------------- \| ------------ \| --------------- \| -------------- \| ------- \| -------- \|
	\| All three \| stages \| are \| realized \| through \| a single \| \| \| \| \| \| \|
	\| \| \| \| \| \| \| and items. \| These \| cues act \| as lightweight \| \| seeds, \|
	\| passinputandoutput: \| \| \| cueextraction,self-prompt \| \| \| \| \| \| \| \| \|
	\| \| \| \| \| \| \| which are \| deliberately \| underspecified \| \| so \| that the \|
	construction,andprofilegenerationareproduced
	systemcansubsequentlyexplorericherprofilefor-
	inasinglesequence-to-sequencegenerationpass
	mats.
	atinferencetime,enablingefficientdeployment.
	\| \| \| \| \| \| \| Definition1(Cue). \| \| Acueisaminimaltextualhy- \| \| \| \|
	\| --- \| --- \| --- \| --- \| --- \| --- \| ----------------- \| --- \| ------------------------ \| --- \| --- \| --- \|
	3.1 ProblemFormulation
	pothesisderivedfromhistoricaldatathathighlights
	WeformulateprofilegenerationinDUETasanon- one potential aspect of a user’s preference or an
	policyreinforcementlearningproblem,motivated item’scharacteristic. Ratherthanaimingforcom-
	bytheabsenceofanytextualgroundtruthdefining pleteness, cues capture partial but salient signals
	an optimal user or item profile. Profile quality is thatserveasstartingpointsforprofileexploration.
	evaluatedsolelybyitsfunctionalutilityinafixed
	\| \| \| \| \| \| \| To extract \| cues \| automatically, \| \| the \| LLM is \|
	\| --- \| --- \| --- \| --- \| --- \| --- \| ---------- \| ---- \| -------------- \| --- \| --- \| ------ \|
	recommendationenvironment.
	\| \| \| \| \| \| \| prompted \| to summarize \| minimal \| \| but informative \| \|
	\| ---- \| ---------- \| --- \| ------------ \| ------ \| ------- \| ------------------------ \| ------------ \| ------- \| ----------------- \| --------------- \| --- \|
	\| DUET \| is modeled \| as \| a generative \| policy \| π θ in- \| \| \| \| \| \| \|
	\| \| \| \| \| \| \| aspectsofuseroritemdata. \| \| \| Forexample,givena \| \| \|
	teractingwithafrozendownstreamrecommender.
	user’sinteractionhistory,themodelisguidedwith
	\| For each \| user–item \| pair, \| the \| state is \| defined as \| \| \| \| \| \| \|
	\| -------- \| --------- \| ----- \| --- \| -------- \| ---------- \| --- \| --- \| --- \| --- \| --- \| --- \|
	instructionssuchas:
	\| s = {H \| ,H },whereH \| \| andH \| denotetheuser \| \| \| \| \| \| \| \|
	\| ------ \| ----------- \| --- \| ---- \| ------------- \| --- \| --- \| --- \| --- \| --- \| --- \| --- \|
	\| \| u i \| \| u \| i \| \| \| \| \| \| \| \|
	and item interaction histories. An action corre- CueExtractionPrompt
	spondstoasingle-passjointgeneration
	“Fromthehistorybelow,analyzetheuser’s
	a = {C u ,S u ,P u ,C i ,S i ,P i }, historicalinteractionstounderstandprefer-
	ences,ratingbehavior,reviewsentimentor
	\| where \| (C ,C \| ) are \| cues \| distilled \| from his- \| \| \| \| \| \| \|
	\| ----- \| ----- \| ----- \| ---- \| --------- \| --------- \| --- \| --- \| --- \| --- \| --- \| --- \|
	u i
	\| \| \| \| \| \| \| anyotherdimension. \| \| Keepthedescription \| \| \| \|
	\| -------- \| ------------ \| ----------- \| --- \| ----------- \| ------- \| ------------------ \| --- \| ------------------ \| --- \| --- \| --- \|
	\| tory, (S \| u ,S i ) are \| constructed \| \| user-item’s \| profile \| \| \| \| \| \| \|
	prompts,and(P ,P )arethefinaltextualprofiles. conciseandavoidfullsentences.”
	u i
	\| Thepolicyπ \| \| (a \| s)definesajointdistribution \| \| \| \| \| \| \| \| \| \|
	\| ---------- \| --- \| -------------------------------- \| --- \| --- \| --- \| --- \| --- \| --- \| --- \| --- \| --- \|
	θ
	over the entire generation sequence. As an on- This lightweight guidance allows the LLM to
	policy agent, DUET is optimized using rewards maprawhistoriesandmetadataintocompacttex-
	from its own sampled generations rather than by tualcues. Thedetaildexampleofcuecanbefound
	\| imitatingfixedsummaries. \| \| \| \| \| \| inAppendixA.3. \| \| \| \| \| \|
	\| ------------------------ \| --- \| --- \| --- \| --- \| --- \| -------------- \| --- \| --- \| --- \| --- \| --- \|

	\| 3.3 \| JointExplorationviaAdaptiveProfile \| \| \| \| \| \| Single-PassInput \| \| \| \| \| \|
	\| --- \| ---------------------------------- \| --- \| --- \| --- \| --- \| --- \| ---------------- \| --- \| --- \| --- \| --- \| --- \|
	PromptDiscovery Task:Generatestructuredprofilesfortheuseranditemfrom
	historicaldata.
	Aftercueextraction,DUETdoesnotdirectlysum- Combined Context: [User History H u ] + [Item
	\| \| \| \| \| \| \| \| History \| H i ]+[Avg \| Ratings] \| \| \| \|
	\| ------------------------------------ \| --- \| --- \| --- \| --- \| --- \| ---------- \| ------- \| ---------- \| -------- \| --- \| --- \| --- \|
	\| marizeuserinterestsoritemattributes. \| \| \| \| \| \| Instead,it \| \| \| \| \| \| \|
	“First,extractaconciseCueforbothuseranditem.Second,
	\| explores \| \| the space \| of profile \| construction \| \| strate- \| \| \| \| \| \| \|
	\| -------- \| --- \| --------- \| ---------- \| ------------ \| --- \| ------- \| --- \| --- \| --- \| --- \| --- \| --- \|
	basedonthecues,constructaProfilePromptthatdefines
	gies—natural-languagepromptsthatdefinethefor- thedescriptionlogic.Finally,generatetheProfileguidedby
	theconstructedProfilePrompt.”
	mat,abstractionlevel,andattributeselectionlogic
	↓DUETForwardPass↓
	\| usedtogenerateprofiles. \| \| \| \| Explorationistherefore \| \| \| \| \| \| \| \| \|
	\| -------------------------------------------- \| --- \| --- \| --- \| ---------------------- \| --- \| --- \| ----------------- \| --- \| --- \| --- \| --- \| --- \|
	\| performedoverhowprofilesshouldbeconstructed, \| \| \| \| \| \| \| Single-PassOutput \| \| \| \| \| \|
	User
	ratherthanoversuperficialtextualparaphrases.
	Cue:“prefersnostalgiclogic-basedchallenges”
	For each user–item pair, DUET intro- Profileprompt: “Focuson1990svisualaestheticsand
	\| duces \| an \| explicit \| intermediate \| \| variable, \| the \| strategicdepth.” \| \| \| \| \| \|
	\| ----- \| --- \| -------- \| ------------ \| --- \| --------- \| --- \| ---------------- \| --------- \| --- \| ----- \| ----------- \| ------------ \|
	\| \| \| \| \| \| \| \| Profile: \| “A player \| who \| seeks \| retro-style \| visual charm \|
	constructed_prompt,whichservesasadiscrete
	pairedwithdeepstrategicreasoning.”
	\| and \| interpretable \| \| profile \| prompt \| S. \| This profile \| \| \| \| \| \| \|
	\| --- \| ------------- \| --- \| ------- \| ------ \| --- \| ------------ \| --- \| --- \| --- \| --- \| --- \| --- \|
	Item
	promptisashortnatural-languageinstruction(e.g.,
	Cue:“retro-styleindiepuzzlewithhighdifficulty”
	“Describetheuser’stypicalgamingpreferencesand
	\| \| \| \| \| \| \| \| Profile \| prompt: \| “Describe \| \| pixel-art \| graphics and \|
	\| --- \| --- \| --- \| --- \| --- \| --- \| --- \| ------- \| ------- \| --------- \| --- \| --------- \| ------------ \|
	engagementpatterns”)thatspecifieswhataspects intellectualdifficultytomatchlogicpreference.”
	Profile:“A2Dexperiencefeaturingpixelatednostalgia
	\| to \| describe \| and \| how to \| organize \| them, \| without \| \| \| \| \| \| \|
	\| --- \| -------- \| --- \| ------ \| -------- \| ----- \| ------- \| --- \| --- \| --- \| --- \| --- \| --- \|
	andchallengingmechanicsthatdemandlogicaldeduction.”
	\| containingtheprofilecontentitself. \| \| \| \| \| \| Conditioned \| \| \| \| \| \| \|
	\| ---------------------------------- \| --- \| --- \| --- \| --- \| --- \| ----------- \| --- \| --- \| --- \| --- \| --- \| --- \|
	onS,themodelgeneratesthefinaluseranditem Figure3: Single-passgenerationinDUET: cueextrac-
	\| profiles. \| \| \| \| \| \| \| tion,profileprompt(constructedprompt),andprofile \| \| \| \| \| \|
	\| --------- \| --- \| --- \| --- \| --- \| --- \| --- \| ------------------------------------------------ \| --- \| --- \| --- \| --- \| --- \|
	generationareproducedinonepassforbothuserand
	\| Exploration \| \| \| is driven \| by treating \| \| the profile \| \| \| \| \| \| \|
	\| ----------- \| --- \| --- \| --------- \| ----------- \| --- \| ----------- \| --- \| --- \| --- \| --- \| --- \| --- \|
	item.
	\| promptS \| \| asastochasticactionsampledfromthe \| \| \| \| \| \| \| \| \| \| \|
	\| ------- \| --- \| --------------------------------- \| --------------------------- \| --- \| --- \| --- \| --- \| --- \| --- \| --- \| --- \| --- \|
	\| policyπ \| \| (S \| Cue). \| Duringtraining,theagentsam- \| \| \| \| \| \| \| \| \| \|
	θ
	plesdifferentprofilepromptinstructions,executes
	\| \| \| \| \| \| \| \| andintellectualdifficulty”tomatch. \| \| \| \| \| RLreinforces \|
	\| ---------------------------- \| --- \| --- \| --- \| --- \| ---------------- \| --- \| ---------------------------------- \| --- \| --- \| --- \| --- \| ------------ \|
	\| themtogenerateprofilepairs(P \| \| \| \| \| ,P ),andreceives \| \| \| \| \| \| \| \|
	u i
	thissharedsemanticdirection,suppressingirrele-
	\| a reward \| \| based \| on downstream \| \| recommendation \| \| \| \| \| \| \| \|
	\| --------- \| --- \| ----------------------------------- \| ------------- \| --- \| -------------- \| --- \| ------------ \| ------------- \| -------- \| --- \| ----- \| ---------------- \|
	\| \| \| \| \| \| \| \| vant signals \| and \| forcing \| the \| final \| profiles to con- \|
	\| accuracy. \| \| Thepolicyisoptimizedtoreinforcepro- \| \| \| \| \| \| \| \| \| \| \|
	\| \| \| \| \| \| \| \| verge \| into a shared \| semantic \| \| space \| of nostalgia \|
	filepromptsthatconsistentlyyieldhigherrewards,
	andlogic,significantlyimprovingrecommendation
	enablingthemodeltoactivelysearchforeffective
	accuracy.
	profileconstructionformats.
	Thisprocessisrealizedasaunifiedgeneration
	3.4 OptimizationviaOn-policyExploration
	pass,
	\| \| \| \| \| \| \| \| On-policy \| optimization. \| \| \| We train \| the profile \|
	\| --- \| --- \| --- \| --- \| --- \| --- \| --- \| --------- \| ------------- \| --- \| --- \| -------- \| ----------- \|
	O = [ Cue → profilepromptS → Profile], generator π in an on-policy manner against a
	\| \| (cid:124)(cid:123)(cid:122)(cid:125) \| \| \| \| \| (cid:124) (cid:123)(cid:122) (cid:125) \| \| θ \| \| \| \| \|
	\| --- \| ------------------------------------ \| --- \| --------- \| ------------------ \| --------- \| -------------------------------------- \| --- \| --- \| --- \| --- \| --- \| --- \|
	\| \| \| \| (cid:124) \| (cid:123)(cid:122) \| (cid:125) \| \| \| \| \| \| \| \|
	Context ProfileConstructionPrompt Execution frozendownstreammodelf, whichservesasthe
	\| \| \| \| \| \| \| (1) \| environment \| critic. \| For \| each \| sampled \| user–item \|
	\| --- \| --- \| --- \| --- \| --- \| --- \| --- \| ----------- \| ------- \| --- \| ---- \| ------- \| --------- \|
	wheretheprofilepromptS parameterizesthepro- pair (u,i), the policy generates (P ,P ) and re-
	u i
	filegenerator. Atinferencetime,DUETcollapses ceivesascalarrewardbasedonthepredictionac-
	thelearnedprofilepromptdistributionandgreed- curacy of f(P ,P ). The policy parameters are
	u i
	ilyexecutestheoptimalprofilepromptinasingle updatedtoreinforcegenerationsthatleadtolower
	\| forwardpass,introducingnoadditionallatency. \| \| \| \| \| \| \| predictionerror. \| \| \| \| \| \|
	\| ------------------------------------------- \| --- \| ---------- \| ------ \| ------------ \| -------- \| ------- \| --------------------------- \| --- \| --- \| --- \| --- \| ------------- \|
	\| Example \| \| of profile \| prompt \| Exploration. \| \| As \| \| \| \| \| \| \|
	\| \| \| \| \| \| \| \| Continuousfractionalreward. \| \| \| \| \| Usingdiscrete \|
	\| illustrated \| \| in Figure \| 3, \| DUET \| resolves \| this by \| \| \| \| \| \| \|
	integerratingsasrewardsleadstosparseandunsta-
	\| identifying \| \| a shared \| context \| in \| a single \| forward \| \| \| \| \| \| \|
	\| ----------- \| --- \| -------- \| ------- \| --- \| -------- \| ------- \| --- \| --- \| --- \| --- \| --- \| --- \|
	blefeedback,asnear-misspredictions(e.g.,predict-
	pass. ThepolicyextractsacoordinatedCue(e.g.,
	ing4foragroundtruth5)receivethesamepenalty
	“prefersretropuzzlegames”)andconstructsapro-
	\| \| \| \| \| \| \| \| assevereerrors. \| \| Toprovidedenseandinformative \| \| \| \|
	\| ---- \| ------ \| ---- \| ------------------- \| --- \| --- \| -------- \| --------------- \| --- \| ---------------------------- \| --- \| --- \| --- \|
	\| file \| prompt \| (the \| constructed_prompt) \| \| \| that in- \| \| \| \| \| \| \|
	feedback,wedefineacontinuousfractionalreward:
	structsthemodeltohighlight“nostalgicaesthetics
	\| and \| strategic \| logic” \| for \| the user, \| while \| simulta- \| \| \| \| \| \| \|
	\| ------- \| ---------- \| ------ \| ---------- \| ---------- \| ----- \| -------- \| --- \| ------- \| --- \| ---- \| ------ \| -------- \|
	\| \| \| \| \| \| \| \| \| \| \| \| \|y −yˆ \| \| \|
	\| \| \| \| \| \| \| \| \| R (u,i) \| \| = 1− \| ui \| ui , (2) \|
	\| neously \| describing \| \| the item’s \| “pixel-art \| \| graphics \| \| perf \| \| \| \| \|
	M

	where yˆ ui = f(P u ,P i ) is the predicted relevance ingsettingbasedondownstreampredictedrating
	score and M is the maximum rating gap (e.g., scores. For each observed user–item interaction,
	M = 4 for a 1–5 scale). This reward provides we randomly sample nine items that the user has
	fine-grainedgradientsthatencourageincremental not interacted with to form a candidate set of ten
	improvementsinrecommendationaccuracy. items, which are ranked in descending order ac-
	We adopt Group Relative Policy Optimization cordingtotheirpredictedratings. Theground-truth
	(GRPO) (DeepSeek-AI et al., 2025) to optimize interacted item is treated as the only positive in-
	theprofilegeneratorundertheabovereward. The stance,andweadoptNDCG@Kastheevaluation
	downstreamrecommenderf isfrozenthroughout metric, with K set to 1, 5, and 10, to assess how
	training, which yields a stable policy optimiza- effectivelythelearnedprofilessupportcorrectitem
	\| tionsettingandpreventsrewarddriftorfeedback- \| \| \| \| ranking. \| \| \| \| \|
	\| -------------------------------------------- \| --- \| --- \| --- \| --------- \| ---------- \| ---------- \| ---- \| ------- \|
	\| \| \| \| \| Baselines \| We compare \| our method \| with \| several \|
	inducedrepresentationcollapse.
	\| \| \| \| \| representative \| baselines. \| 10H \| directly \| uses the \|
	\| --- \| --- \| --- \| --- \| -------------- \| ---------- \| --- \| -------- \| -------- \|
	4 Experiment mostrecentinteractionhistoriesforpredictionwith-
	\| \| \| \| \| outconstructingexplicitprofiles. \| \| \| KAR(Xietal., \| \|
	\| --- \| --- \| --- \| --- \| -------------------------------- \| --- \| --- \| ------------ \| --- \|
	4.1 ExperimentalSettings
	2024)augmentsrecommendationmodelswithex-
	Datasets. Experiments were conducted on three ternalreasoningknowledgeaboutuserpreferences
	widelyusedreal-worlddatasets. AmazonMusic andfactualknowledgeaboutitemsextractedfrom
	(Music) and Amazon Book (Book) are derived LLMs,whicharetransformedintotask-compatible
	fromtheAmazonProductdataset1,whileYelpis representations. RLMRec(Renetal.,2024)lever-
	fromtheYelpOpendataset2. Alldatasetsinclude agesLLMstolearnsemanticuseranditemrepre-
	userreviews,ratings,andrichtextualinformation. sentationsfromtextualsignalsandalignsthemwith
	WeusedthefullAmazonMusicdataset,butonly collaborativerelationalinformationthroughcross-
	subsets (latest two months for Book and six for viewrepresentationlearning. PALR(Yangetal.,
	Yelp). Data was split by timestamp into training, 2023)fine-tunesalargelanguagemodelasarank-
	validation,andtestsetstopreventinformationleak- ing component that selects preferred items from
	age(Jietal.,2023). retrievedcandidatesexpressedinnaturallanguage.
	EvaluationMetrics LG (LettinGo) (Wang et al., 2025) explores di-
	For each observed user–item interaction (i.e., verseuserprofilecandidateswithLLMsandaligns
	a review record), we construct the evaluation profilegenerationwithdownstreamrecommenda-
	instance based on the user’s interaction history tionperformanceviapreferenceoptimization. Rea-
	strictly prior to the corresponding timestamp. son4Rec(Fangetal.,2025)introducesadelibera-
	tiverecommendationframeworkthatincorporates
	Specifically,theuser’srecenthistoricalinteractions
	beforethecurrentinteractionareusedtogenerate explicitstep-wisereasoningoveruserpreferences
	auserprofile,whilehistoricalreviewsfromother toguideratingprediction.
	\| users are used \| to generate \| an item profile. \| The \| \| \| \| \| \|
	\| -------------- \| ----------- \| ---------------- \| --- \| --- \| --- \| --- \| --- \| --- \|
	4.2 MainResults
	downstreamrecommendationsystemthenpredicts
	theratingofthetargetitemconditionedontwoin- Table 1 presents a comparison of our proposed
	puts: thegenerateduserprofile,andthegenerated method against five baselines on three datasets:
	\| \| \| \| \| Amazon \| Music, Amazon \| Books, \| and \| Yelp. We \|
	\| ------------ \| ------------------------------- \| --- \| --- \| ------ \| ------------- \| ------ \| --- \| -------- \|
	\| itemprofile. \| Anexceptionisthe10Hbaseline,for \| \| \| \| \| \| \| \|
	which the downstream model directly consumes use Qwen3-8B (Team, 2025) and LLaMA3-8B
	therawrecentinteractionhistories. (Dubey et al., 2024) as both the profile generator
	andthedownstreamrecommendationmodel,with
	\| We evaluate \| performance \| using four \| widely \| \| \| \| \| \|
	\| --------------- \| ----------- \| ------------------ \| ------- \| ------------ \| ----------- \| ----- \| ----------- \| ------- \|
	\| \| \| \| \| a prediction \| temperature \| of 0. \| In addition \| to rat- \|
	\| adopted metrics \| (Wang \| et al., 2025; Fang \| et al., \| \| \| \| \| \|
	2025): MeanAbsoluteError(MAE),RootMean ingpredictionresults,wealsoreportranking-based
	SquareError(RMSE),Accuracy,andF1score. evaluation results in Table ??, which assess the
	effectivenessofthelearnedprofilesfromadown-
	Inadditiontoratingprediction,wefurtherevaluate
	thegenerateduseranditemprofilesunderarank- streamrankingperspective.
	\| \| \| \| \| Overallsuperiorityoverstrongbaselines. \| \| \| \| As \|
	\| --- \| --- \| --- \| --- \| -------------------------------------- \| --- \| --- \| --- \| --- \|
	1https://cseweb.ucsd.edu/~jmcauley/datasets/ shown in Table 1, our method consistently out-
	amazon/links.html
	\| \| \| \| \| performs \| the strongest \| baselines \| across \| all three \|
	\| --- \| --- \| --- \| --- \| -------- \| ------------- \| --------- \| ------ \| --------- \|
	2https://business.yelp.com/data/resources/
	\| open-dataset/ \| \| \| \| datasetsunderbothQwen3-8BandLLaMA3-8B. \| \| \| \| \|
	\| ------------- \| --- \| --- \| --- \| -------------------------------------- \| --- \| --- \| --- \| --- \|

	Method Yelp AmazonMusic AmazonBooks
	MAE RMSE Acc F1(%) MAE RMSE Acc F1(%) MAE RMSE Acc F1(%)
	(%) (%) (%)
	Qwen3(8B)
	10H 1.1235 1.9478 23.17 27.54 0.9102 1.4021 39.26 46.58 0.9314 1.4527 37.63 45.19
	KAR(Xietal.,2024) 0.7396 1.2184 55.34 48.67 0.7483 1.1380 58.65 60.29 0.7098 1.0923 56.17 58.78
	RLMRec(Renetal.,2024) 0.8197 1.3312 47.15 42.46 0.7438 1.1069 54.89 57.65 0.7812 1.1584 52.86 55.93
	PALR(Yangetal.,2023) 0.7994 1.2876 48.53 43.19 0.6075 0.9531 57.35 56.77 0.7485 1.1187 54.24 56.38
	LG(Wangetal.,2025) 0.6632 1.1047 56.18 48.95 0.4737 0.8834 62.37 57.09 0.5821 0.9416 59.35 60.57
	R4Rec(Fangetal.,2025) 0.7028 1.1523 55.69 47.73 0.5654 0.9635 58.69 54.67 0.6397 1.0098 58.47 56.84
	Ours 0.5126 0.9485 61.23 55.18 0.3937 0.7564 67.96 63.89 0.4612 0.9089 64.38 59.27
	LLaMA3(8B)
	10H 1.0864 1.9532 22.09 27.30 0.7917 1.3346 38.13 46.87 0.8064 1.3866 37.15 45.27
	KAR(Xietal.,2024) 0.6427 1.1668 54.51 47.98 0.5726 0.9033 57.53 59.92 0.5892 0.9614 55.87 58.21
	RLMRec(Renetal.,2024) 0.7428 1.3572 46.74 42.11 0.6076 0.9886 53.78 57.42 0.6226 0.9477 52.12 55.79
	PALR(Yangetal.,2023) 0.7238 1.3265 47.72 43.29 0.5823 0.9222 56.73 59.31 0.5977 0.8855 55.06 57.62
	LG(Wangetal.,2025) 0.6196 1.1289 56.03 51.24 0.5204 0.9369 61.92 59.50 0.5543 0.7967 58.95 60.39
	R4Rec(Fangetal.,2025) 0.7586 1.0418 55.80 53.00 0.5442 0.7722 60.86 54.88 0.6029 0.8345 59.70 56.35
	Ours 0.5367 0.9687 60.87 54.74 0.4680 0.8277 63.30 60.60 0.5092 0.9500 63.42 58.12
	Table1: PerformanceonthreedatasetsusingQwen3(8B)andLLaMA3(8B).
	Method Yelp AmazonMusic AmazonBooks
	NDCG@1 NDCG@5 NDCG@10 NDCG@1 NDCG@5 NDCG@10 NDCG@1 NDCG@5 NDCG@10
	10H 0.1823 0.2815 0.4928 0.1875 0.3796 0.5153 0.1841 0.3146 0.4263
	KAR(Xietal.,2024) 0.2156 0.3298 0.5412 0.3018 0.4896 0.6015 0.2965 0.4715 0.5834
	RLMRec(Renetal.,2024) 0.2419 0.3472 0.5587 0.3371 0.5434 0.6162 0.2748 0.4526 0.5719
	PALR(Yangetal.,2023) 0.2494 0.3563 0.5691 0.3395 0.5247 0.6115 0.2627 0.4634 0.5538
	LG(Wangetal.,2025) 0.3187 0.4685 0.5814 0.4012 0.5674 0.6489 0.3795 0.5189 0.6284
	R4Rec(Fangetal.,2025) 0.2575 0.3792 0.5526 0.2928 0.5912 0.6343 0.3013 0.4928 0.5959
	Ours 0.3390 0.4873 0.6008 0.5123 0.6165 0.7025 0.4288 0.5638 0.6599
	Table2: RankingperformanceunderEASE-based(Steck,2019)hardnegatives.
	Under Qwen3-8B, our method achieves an accu- our approach reaches an NDCG@1 of 0.5619
	racyof61.23%onYelp,67.96%onAmazonMu- and an NDCG@10 of 0.7443, clearly outper-
	sicand64.38%onAmazonBooks,surpassingLG forming the strongest baseline methods. Similar
	(LettinGo)(Wangetal.,2025)by5.05%,5.59% trendsareobservedonAmazonMusic,whereour
	and5.03%,respectively. Similarimprovementsare methodattains0.5347atNDCG@1and0.7331at
	observed under LLaMA3-8B, indicating that the NDCG@10, as well as on Amazon Books with
	gainsarestableacrossbackbonemodels. NDCG@1 of 0.4866 and NDCG@10 of 0.7107.
	Theseresultsindicatethatthelearnedprofilesen-
	Advantages over fixed-structure and delib-
	able more accurate identification and ordering of
	erative baselines. Compared with strong fixed-
	relevantitemswithincandidatesets,leadingtosu-
	structureordeliberativemethodssuchasKAR(Xi
	perior ranking quality in downstream recommen-
	etal.,2024)andReason4Rec(Fangetal.,2025),
	dation. Wefurtherevaluaterankingperformance
	our approach consistently achieves lower predic-
	under a more challenging setting using EASE-
	tion error and higher accuracy. For example, on
	based (Steck, 2019) hard negatives, and observe
	Amazon Books with Qwen3-8B, our method re-
	consistentimprovements;detailedresultsarepro-
	duces MAE to 0.4612, compared to 0.7098 for
	videdinAppendixB.2.
	KARand0.6397forReason4Rec,demonstrating
	moreeffectiveabstractionofuserpreferenceswith-
	4.3 AblationStudy
	outrelyingonpredefinedreasoningtemplates.
	Consistent improvements in downstream Effectivenessofprofilegeneration,cue&strategy,
	ranking. Table??reportstherankingresultsun- anduser–itemjointoptimization. Table3reports
	der the Qwen3-8B backbone. Our method con- the ablation results under different design con-
	sistentlyachievesthebestperformanceacrossall figurations. Starting from the history-only base-
	three datasets and evaluation cutoffs. On Yelp, line (10H), introducing explicit profile genera-

	\| Method \| \| \| \| Yelp \| \| \| AmazonMusic \| \| \| AmazonBooks \| \|
	\| ------ \| --- \| --- \| --- \| ---- \| --- \| ------ \| ----------- \| --- \| --- \| ----------- \| ------ \|
	\| \| \| \| MAE \| RMSE \| Acc \| F1 MAE \| RMSE \| Acc \| F1 \| MAE RMSE \| Acc F1 \|
	10H(HistoryOnly) 1.1235 1.9478 23.17 27.54 0.9102 1.4021 39.26 46.58 0.9314 1.4527 37.63 45.19
	Profile 0.7218 1.1863 55.48 48.09 0.6597 1.0218 58.67 57.48 0.6764 1.0469 57.14 57.68
	Profile+Cue&Strategy 0.7085 1.1654 55.83 48.54 0.5708 0.9897 58.91 55.53 0.6389 1.0108 58.43 56.88
	Profile+JointOpt.(LG(Wang 0.6632 1.1047 56.18 48.95 0.4737 0.8834 62.37 57.09 0.5821 0.9416 59.35 60.57
	etal.,2025))
	Profile+Cue&Strategy+ 0.5126 0.9485 61.23 55.18 0.3937 0.7564 67.96 63.89 0.4612 0.9089 64.38 59.27
	JointOpt.(Ours)
	Table3: AblationstudyondifferentdesignconfigurationsinDUETusingQwen3(8B).
	\| Method \| \| Yelp \| \| \| \| AmazonMusic \| \| \| \| AmazonBooks \| \|
	\| ------ \| -------- \| ---- \| --- \| ----- \| --- \| ----------- \| --- \| ----- \| --- \| ----------- \| --------- \|
	\| \| MAE RMSE \| \| Acc \| F1(%) \| MAE \| RMSE \| Acc \| F1(%) \| MAE \| RMSE \| Acc F1(%) \|
	\| \| \| \| (%) \| \| \| \| (%) \| \| \| \| (%) \|
	10H+30P 0.5126 0.9485 61.23 55.18 0.3883 0.7494 67.96 63.89 0.4612 0.9089 65.13 59.97
	10H+50P 0.4909 0.9207 62.43 56.24 0.3924 0.7543 67.88 63.87 0.4553 0.9023 64.62 59.52
	10H+70P 0.4987 0.9326 61.98 55.81 0.3937 0.7564 68.22 64.12 0.4608 0.9068 64.38 59.27
	Table4: Impactofhistoricalinteractionlengthonprofilequality(usingQwen3(8B)).
	tionyieldssubstantialperformanceimprovements jointoptimizationarecombined. Underthisfull
	across all datasets. For example, on Yelp, MAE configuration, accuracy reaches 61.23% on Yelp,
	isreducedfrom1.1235to0.7218andaccuracyin- 67.96%onAmazonMusic,and64.38%onAma-
	creases from 23.17% to 55.48%, confirming that zon Books, with corresponding MAE values of
	textualprofilesprovidesignificantlymoreinforma- 0.5126, 0.3937, and 0.4612, respectively. Com-
	tiverepresentationsthanrawinteractionhistories. pared with the LettinGo-style configuration, this
	\| \| \| \| \| \| \| setting \| yields \| consistent \| \| improvements \| across all \|
	\| ------ \| ---------------- \| --- \| ----- \| --- \| ------ \| --------- \| ------ \| -------------------------------------- \| --- \| ------------ \| ---------- \|
	\| Adding \| the cue&strategy \| \| layer \| on \| top of \| pro- \| \| \| \| \| \|
	\| \| \| \| \| \| \| datasets. \| \| Overall,theablationresultsindicatethat \| \| \| \|
	filegenerationleadstomodestbutconsistentgains.
	whileprofilegenerationandjointoptimization
	OnAmazonMusic,MAEfurtherdecreasesfrom
	0.6597to0.5708,whileaccuracyslightlyimproves contributesubstantiallytoperformancegains,inte-
	gratingcue&strategydiscoveryfurtherenhances
	\| from58.67%to58.91%. \| \| \| Althoughthenumerical \| \| \| \| \| \| \| \| \|
	\| ------------------- \| --- \| --- \| -------------------- \| --- \| --- \| --- \| ------------- \| --- \| -------- \| --------- \| ------------ \|
	\| \| \| \| \| \| \| the \| effectiveness \| \| of joint \| user–item \| optimization \|
	improvementsintroducedbystrategyalonearelim-
	ited, these results suggest that strategy discovery byprovidingastructuredspaceforexplorationand
	refinement.
	primarilyreshapeshowpreferenceinformationis
	abstractedandexpressed,ratherthandirectlyopti- Impactofhistoricalinteractionlengthonprofile
	mizingpredictionaccuracyinisolation.
	\| \| \| \| \| \| \| quality. \| \| Table \| 4 analyzes \| the \| impact of histori- \|
	\| --- \| --- \| --- \| --- \| --- \| --- \| -------- \| --- \| ----- \| ---------- \| --- \| ------------------ \|
	Whenenablingjointoptimizationwithoutstrat- cal interaction length on profile quality. Across
	thethreedatasets,varyingthenumberofhistorical
	\| egy (i.e., \| the LettinGo \| (Wang \| et \| al., 2025)-style \| \| \| \| \| \| \| \|
	\| ---------- \| ------------ \| ----- \| --- \| ---------------- \| --- \| --- \| --- \| --- \| --- \| --- \| --- \|
	configuration),performanceimprovesmorenotice- interactions leads to only moderate performance
	ablyacrossdatasets. OnAmazonMusic,accuracy differences, indicating that our method does not
	\| \| \| \| \| \| \| strongly \| \| depend \| on long \| histories. \| In particular, \|
	\| --- \| --- \| --- \| --- \| --- \| --- \| -------- \| --- \| ------ \| ------- \| ---------- \| -------------- \|
	increasesto62.37%,comparedto58.67%withpro-
	filegenerationalone. Inourimplementation,this usingamoderatehistorylength(e.g.,30–50inter-
	settingcorrespondstoareproductionofLettinGo, actions)alreadyachievescompetitiveorbestper-
	whereprofilegenerationandjointoptimizationare formanceonmostmetrics,whilefurtherincreasing
	thehistorylengthprovideslimitedadditionalgains.
	\| appliedwithoutthestrategylayer. \| \| \| \| Whiletheorig- \| \| \| \| \| \| \| \|
	\| ------------------------------- \| ------- \| --------- \| --- \| ------------- \| ---------- \| --- \| --- \| --- \| --- \| --- \| --- \|
	\| inal method \| focuses \| primarily \| on \| user \| profiling, \| \| \| \| \| \| \|
	OnYelpandAmazonBooks,extendingthehis-
	weextendtheoptimizationtobothuseranditem
	torylengthbeyondthisrangedoesnotconsistently
	profilestoensureafaircomparison. improveaccuracyorF1scoreandmayslightlyde-
	Thestrongestandmostconsistentperformance gradeperformance,suggestingthatexcessivehis-
	is achieved when cue&strategy discovery and toricalinteractionscanintroducenoisyorlessrele-

	vantsignals. Incontrast,AmazonMusicexhibits maintaining comparable coverage. Compared to
	relativelystableperformanceacrossdifferenthis- existing methods, which either rely on extractive
	torylengths,indicatingthatuserpreferencesinthis compression(e.g.,KAR(Xietal.,2024))orgener-
	domainarelesssensitivetohistorytruncation. atefree-formdescriptionswithweakergrounding
	Necessity of RL-based Optimization. We com- (e.g.,RLMRec(Renetal.,2024)andPALR(Yang
	pare the full model with a variant without RL etal.,2023)), DUET consistentlyattainsstronger
	(DUET W/O RL) to isolate the effect of RL. As alignment without sacrificing coverage. For in-
	showninTable5,removingRLleadstosubstantial stance,LG(Wangetal.,2025)achievesrelatively
	performancedegradationacrossalldatasets(e.g., highalignmentbutexhibitslessstablegrounding
	YelpAccuracydropsfrom61.23%to48.53%),indi- acrossdatasets,whileRLMRecattainshighercov-
	catingthatthegainscannotbeattributedtoprompt erageatthecostofweakersemanticalignment. In
	\| designalone. \| \| \| \| \| contrast,DUETmaintainsbothhighalignmentand \| \| \| \| \| \|
	\| ------------ \| --- \| --- \| --- \| --- \| ------------------------------------------ \| --- \| --- \| --- \| --- \| --- \|
	Without RL, the generator reduces to a static mid-to-highcoverage,suggestingamoreeffective
	mapping from interaction history to textual pro- balancebetweensemanticabstractionandevidence
	\| files,lackingadaptiveselectionofrelevantsignals. \| \| \| \| \| preservation. \| \| \| \| \| \|
	\| ------------------------------------------------ \| --- \| --- \| --- \| --- \| ------------- \| --- \| --- \| --- \| --- \| --- \|
	Incontrast,RLenablesoptimizingprofileconstruc-
	4.5 CaseStudy
	tionunderrewardfeedback,resultinginmoredis-
	Figure4presentsarepresentativecasestudythat
	\| criminative \| representations.These \| \| results \| demon- \| \| \| \| \| \| \|
	\| ----------- \| --------------------- \| --- \| ------- \| ------ \| --- \| --- \| --- \| --- \| --- \| --- \|
	strate that RL-based optimization is essential for demonstrates how semantically aligned user and
	\| DUET \| . \| \| \| \| itemprofiles \| enableaccurate \| \| prediction \| \| thattran- \|
	\| ---- \| --- \| --- \| --- \| --- \| ------------ \| -------------- \| --- \| ---------- \| --- \| --------- \|
	scendsthelimitationsofsparserawinteractionhis-
	4.4 SemanticAnalysisofGeneratedProfiles tories. In the user profile, initially scattered and
	To better understand the source of performance fragmentedpreferencecuesaresystematicallydis-
	gains, we analyze whether the generated profiles tilled into a stable and coherent preference struc-
	exhibitmeaningfulsemanticpropertiesratherthan turefavoringfunk,soul,andprogressiverock(or-
	servingasintermediatetextualartifacts. Weintro- ange), coupled with a pronounced emphasis on
	duce two complementary metrics to characterize musicalcomplexityandhistoricalsignificance(pur-
	semanticcompatibilityandgrounding. ple). Thecorrespondingitemprofileexhibitsstrik-
	Semantic Alignment. We measure the ingsymmetry,characterizingthealbumasanexem-
	embedding-levelsimilaritybetweengenerateduser plaryfunk-rockcomposition(orange)andposition-
	anditemprofilesusingall-mpnet-base-v2from ingitasadefiningandculturallyinfluentialrelease
	Sentence-Transformers (Reimers and Gurevych, of the 1970s (purple). This example shows 519
	2019). Foreachuser–itempair,wecomputecosine that the learned profiles capture meaningful pref-
	\| similarity: \| \| \| \| \| erence–attributecorrespondencethatisdifficultto \| \| \| \| \| \|
	\| -------------------------------- \| -------------------------------- \| --- \| ------------ \| --- \| ----------------------------------------------- \| --------- \| ------- \| ---------- \| ---- \| ----------- \|
	\| \| \| \| e ·e \| \| recoverfromindividualreviewsalone. \| \| \| \| \| \|
	\| \| \| \| u i \| \| \| \| \| \| \| \|
	\| \| Align(u,i) \| = \| \| (3) \| \| \| \| \| \| \|
	\| \| \| \| ∥e ∥∥e ∥ \| \| \| \| \| \| \| \|
	\| \| \| \| u i \| \| 5 Conclusion \| \| \| \| \| \|
	\| wheree \| ande denotetheembeddingvectorsof \| \| \| \| \| \| \| \| \| \|
	\| \| u i \| \| \| \| In this \| paper, we \| propose \| DUET \| , a \| closed-loop \|
	\| thegenerateduseranditemprofiles. \| \| \| Highervalues \| \| \| \| \| \| \| \|
	\| \| \| \| \| \| framework \| for \| jointly \| generating \| user \| and item \|
	indicatestrongersemanticcompatibilitybetween
	\| \| \| \| \| \| textualprofilesforrecommendation. \| \| \| \| \| Unlikeprior \|
	\| --- \| --- \| --- \| --- \| --- \| --------------------------------- \| --- \| --- \| --- \| --- \| ----------- \|
	modeleduserpreferencesanditemcharacteristics.
	\| \| \| \| \| \| methods \| that rely \| on \| fixed templates \| \| or indepen- \|
	\| -------- \| --------------- \| --- \| ---------- \| ------ \| ------- \| ----------- \| --------- \| --------------- \| --- \| -------------- \|
	\| Coverage \| (Faithfulness). \| \| We measure \| token- \| \| \| \| \| \| \|
	\| \| \| \| \| \| dently \| constructed \| profiles, \| DUET \| \| treats profile \|
	levelgroundingas:
	\| \| \| \| \| \| generation \| as an \| exploration \| problem \| \| and aligns \|
	\| --- \| --- \| --- \| --- \| --- \| ---------- \| ----- \| ----------- \| ------- \| --- \| ---------- \|
	\|Tokens(profile)∩Tokens(history)\| representations directly with downstream recom-
	Cov =
	\| \| \| \|Tokens(profile)\| \| \| \| mendationperformance. \| \| \| \| \| \|
	\| --- \| --- \| ----------------- \| --- \| --- \| --------------------- \| --- \| ---- \| ---------- \| --------- \| ---- \|
	\| \| \| \| \| (4) \| Specifically, \| \| DUET \| integrates \| cue-based \| ini- \|
	Thismetricquantifieshowmuchofthegenerated tialization, adaptive strategy construction, and
	profileissupportedbyhistoricaltextualevidence. feedback-drivenjointoptimizationtoproduceflex-
	We report coverage separately for user and item ible yet task-aligned user–item profiles. By opti-
	\| profiles. \| \| \| \| \| mizing \| both profiles \| in \| a shared \| semantic \| space, \|
	\| --------- \| --- \| --- \| --- \| --- \| ------ \| ------------- \| --- \| -------- \| -------- \| ------ \|
	AsshowninTable6,DUETachievesthehigh- theframeworkreducessemanticmismatchandcap-
	est semantic alignment across all datasets while turesinteraction-relevantsignalsthataredifficultto

	Setting Yelp AmazonMusic AmazonBooks
	MAE RMSE Acc F1 MAE RMSE Acc F1 MAE RMSE Acc F1
	(%) (%) (%) (%) (%) (%)
	DUETw/oRL 0.8283 1.3893 48.53 41.15 0.7322 1.1430 57.18 52.48 0.8741 1.4760 51.83 50.35
	DUET(full) 0.5126 0.9485 61.23 55.18 0.3937 0.7564 67.96 63.89 0.4612 0.9089 64.38 59.27
	Table5: EffectofRL-basedoptimization(usingQwen3(8B)).
	User History:
	Andre' S Grindle "Tour of U.S.A.", Rating: 5.0. If your looking for heavy metals dim orgins look no further.
	Andre' S Grindle "Misfits", Rating:4.0. Actually the opening title track is more of a direct pop/rock type song. But the rock n roll
	returns in a big way on "Hay Fever".
	............
	Item History:
	Worlds Away, Rating: 4.0, Review: The CD is wonderful, and is a seminal piece of music for both the group and for those of us who
	were 'of age' in the '70s and '80s. Pablo Cruise was - and is - a group that makes beautiful music like few others, and this
	shipper/seller made it all the more special with a fast ship! THANKS!!!
	Worlds Away, Rating: 5.0, I love this CD! I'd recommend the Pablo Cruise Millenium Collection to get to know this band. Worlds Away to
	me represents the best of breezy '70s pop. Five stars!
	............
	USER PROFILE:
	Andre Grindle is a dedicated music enthusiast with a deep appreciation for funk, soul, and progressive rock, particularly drawn to artists
	who blend technical mastery with genre-defining innovation. His reviews consistently highlight albums that showcase musical complexity,
	lyrical depth, and historical significance, often referencing artists like Stevie Wonder, Prince, and Rush. He values authenticity and
	artistic vision, frequently praising albums that resist trend-chasing and prioritize creative integrity. His detailed, introspective reviews
	suggest a preference for nuanced analysis over superficial praise, with a focus on emotional resonance, production quality, and the
	artist's evolution. While he is generous with high ratings (often 5.0), he is also willing to critique perceived shortcomings, as seen in his
	4.0 for "Test for Echo" and 3.0 for "The Color of Success." His engagement is characterized by a blend of personal anecdotes, technical
	appreciation, and historical context, indicating a preference for music that transcends mere entertainment.
	Item PROFILE:
	"Worlds Away" by Pablo Cruise is celebrated as a peak-era funk-rock album that exemplifies the band's ability to merge polished
	production with genre-blurring creativity. Users consistently praise its tight musicianship, genre-defying sound, and the band's ability to
	craft catchy, emotionally resonant tracks. The album's appeal lies in its seamless fusion of rock, funk, and pop, with standout tracks like
	"Love Will Find a Way" and "Worlds Away" receiving particular acclaim for their infectious melodies and lyrical depth. While some users
	note its dated production compared to modern standards, the majority highlight its historical significance as a defining work of the
	1970s rock-funk movement. The album resonates most with fans of 1970s progressive rock, genre-defying artists, and listeners seeking
	well-crafted, emotionally engaging music. Its high average rating reflects widespread admiration for its artistic ambition and enduring
	appeal, though a minority acknowledges its limitations in terms of modern production quality.
	Figure4: Illustrationofthemutualcorrespondencebetweenuseranditem. Thehighlightedregionsdemonstrate
	thatuserpreferencessummarizedintheuserprofilealignwiththekeyattributesextractedintheitemprofile,which
	providescomplementaryinformationbeyondrawhistoriesandthusimprovespredictionaccuracy.
	Method Yelp AmazonMusic AmazonBooks
	Align UserCov. ItemCov. Align UserCov. ItemCov. Align UserCov. ItemCov.
	10H 0.3902 /(1.00) /(1.00) 0.3577 /(1.00) /(1.00) 0.2604 /(1.00) /(1.00)
	KAR(Xietal.,2024) 0.4932 0.1745 0.1823 0.4807 0.2170 0.2391 0.5702 0.1877 0.1544
	RLMRec(Renetal.,2024) 0.4010 0.2646 0.3773 0.3931 0.3526 0.4889 0.4436 0.2994 0.3506
	PALR(Yangetal.,2023) 0.4715 0.3675 0.1917 0.4938 0.2197 0.2390 0.5216 0.2050 0.2267
	LG(Wangetal.,2025) 0.5709 0.2378 0.2682 0.5109 0.2546 0.3749 0.5506 0.3883 0.3364
	R4Rec(Fangetal.,2025) 0.4882 0.2348 0.2330 0.4208 0.3328 0.3595 0.4946 0.2831 0.2949
	Ours 0.6382 0.2880 0.3429 0.5947 0.4002 0.4482 0.7287 0.3457 0.3127
	Table6: Semanticalignmentandcoverageofgeneratedprofiles. Cov. denotesthefractionoftokensgroundedin
	historicaltext.
	recoverfromrawhistoriesorstaticpromptsalone. reinforcementlearning–basedoptimization. These
	resultssuggestthatadaptive,interaction-awaretex-
	Experiments on multiple real-world datasets tualprofilesprovideapromisingdirectionformore
	demonstratethatDUETconsistentlyoutperforms effective interpretable and performance-oriented
	strongbaselinesunderdifferentbackbonemodels, recommendationsystems.
	validating the effectiveness of joint profiling and

	\| Limitations \| \| \| \| \| \| \| References \| \| \| \| \| \| \|
	\| ----------- \| --- \| --- \| --- \| --- \| --- \| --- \| ---------- \| ----- \| ------ \| ---- \| ------ \| ------ \| ----- \|
	\| \| \| \| \| \| \| \| Keqin Bao, \| Jizhi \| Zhang, \| Yang \| Zhang, \| Wenjie \| Wang, \|
	Despiteitseffectiveness,ourapproachhasseveral FuliFeng,andXiangnanHe.2023. Tallrec: Anef-
	fectiveandefficienttuningframeworktoalignlarge
	\| limitations. \| First, \| the \| proposed \| framework \| \| relies \| \| \| \| \| \| \| \|
	\| ------------ \| ------ \| --- \| -------- \| --------- \| --- \| ------ \| -------------------------------- \| --- \| --- \| --- \| --- \| ---------- \| --- \|
	\| \| \| \| \| \| \| \| languagemodelwithrecommendation. \| \| \| \| \| InProceed- \| \|
	onlargelanguagemodelsforbothprofilegenera- ingsofthe17thACMConferenceonRecommender
	tion and downstream recommendation, which in- Systems,pages1007–1014.
	\| troduces \| additional \| computational \| \| \| overhead \| dur- \| \| \| \| \| \| \| \|
	\| -------- \| ---------- \| ------------- \| --- \| --- \| -------- \| ---- \| --- \| --- \| --- \| --- \| --- \| --- \| --- \|
	TingChen,Wei-LiHan,Hai-DongWang,Yi-XunZhou,
	\| ingtrainingandinference. \| \| \| Whileourexperiments \| \| \| \| \| \| \| \| \| \| \|
	\| ------------------------ \| --- \| --- \| ------------------- \| --- \| --- \| --- \| ------------------------- \| --- \| --- \| --- \| ---------------- \| --- \| --- \|
	\| \| \| \| \| \| \| \| BinXu,andBin-YuZang.2007. \| \| \| \| Contentrecommen- \| \| \|
	showconsistentgainsacrossdifferentbackbones, dationsystembasedonprivatedynamicuserprofile.
	the overall efficiency may be constrained when In2007Internationalconferenceonmachinelearn-
	\| \| \| \| \| \| \| \| ing and \| cybernetics, \| \| volume \| 4, pages \| 2112–2118. \| \|
	\| --------------------------------- \| --- \| --- \| --- \| --- \| ---------- \| --- \| ------- \| ------------ \| --- \| ------ \| -------- \| ---------- \| --- \|
	\| scalingtoverylargeuseroritemsets. \| \| \| \| \| Second,the \| \| \| \| \| \| \| \| \|
	IEEE.
	\| quality \| of the generated \| \| profiles \| is \| inherently \| de- \| \| \| \| \| \| \| \|
	\| ------- \| ----------------- \| --- \| -------- \| --- \| ------------- \| --- \| --------- \| ------- \| --- \| ----------- \| ----- \| --- \| ----- \|
	\| pendent \| on the underlying \| \| LLMs \| \| and prompting \| \| \| \| \| \| \| \| \|
	\| \| \| \| \| \| \| \| Yue Chen, \| Minghua \| \| He, Fangkai \| Yang, \| Pu \| Zhao, \|
	strategies. Variationsinmodelcapacityorprompt LuWang,YuKang,YifeiDong,YuefengZhan,Hao
	Sun,QingweiLin,SaravanRajmohan,andDongmei
	sensitivitymayleadtodifferencesinprofilestabil-
	\| \| \| \| \| \| \| \| Zhang. \| 2025. \| Warriormath: \| \| Enhancing \| the \| mathe- \|
	\| --- \| --- \| --- \| --- \| --- \| --- \| --- \| ------ \| ----- \| ------------ \| --- \| --------- \| --- \| ------ \|
	ity,whichwedonotexplicitlycontrolinthecurrent
	maticalabilityoflargelanguagemodelswithadefect-
	design. Finally,ourevaluationfocusesontext-rich
	\| \| \| \| \| \| \| \| awareframework. \| \| \| Preprint,arXiv:2508.01245. \| \| \| \|
	\| --- \| --- \| --- \| --- \| --- \| --- \| --- \| --------------- \| --- \| --- \| -------------------------- \| --- \| --- \| --- \|
	recommendationscenarioswheresufficienthistor-
	\| \| \| \| \| \| \| \| Paul Covington, \| \| Jay \| Adams, and \| Emre \| Sargin. \| 2016. \|
	\| ------------ \| --- \| ---------- \| --- \| ----------------- \| --- \| --- \| --------------- \| --- \| --- \| ---------- \| ---- \| ------- \| ----- \|
	\| ical reviews \| are \| available. \| \| The effectiveness \| \| of \| \| \| \| \| \| \| \|
	Deepneuralnetworksforyoutuberecommendations.
	\| the proposed \| strategy \| \| in domains \| with \| extremely \| \| \| \| \| \| \| \| \|
	\| ------------ \| -------- \| --- \| ---------- \| ---- \| --------- \| --- \| --- \| --- \| --- \| --- \| --- \| --- \| --- \|
	InProceedingsofthe10thACMconferenceonrec-
	sparsetextualsignalsorinnon-textualmodalities
	ommendersystems,pages191–198.
	remainstobefurtherexplored.
	DeepSeek-AI,DayaGuo,DejianYang,HaoweiZhang,
	JunxiaoSong,RuoyuZhang,RunxinXu,QihaoZhu,
	ShirongMa,PeiyiWang,XiaoBi,XiaokangZhang,
	Data,Privacy,andEthicsConsiderations Xingkai Yu, Yu Wu, Z. F. Wu, Zhibin Gou, Zhi-
	\| \| \| \| \| \| \| \| hong \| Shao, \| Zhuoshu \| Li, Ziyi \| Gao, and \| 181 \| others. \|
	\| --- \| --- \| --- \| --- \| --- \| --- \| --- \| ----- \| ------------ \| ------- \| ------------- \| --------- \| --- \| ------- \|
	\| \| \| \| \| \| \| \| 2025. \| Deepseek-r1: \| \| Incentivizing \| reasoning \| \| capa- \|
	Alldatasetsusedinthisworkarepubliclyreleased bilityinllmsviareinforcementlearning. Preprint,
	arXiv:2501.12948.
	bytheiroriginalowners,andwefollowtheofficial
	termsofuseassociatedwitheachdataset,utilizing
	JiaxinDeng,ShiyaoWang,KuoCai,LejianRen,Qigen
	thedatasolelyforresearchpurposesaspermitted
	\| \| \| \| \| \| \| \| Hu, Weifeng \| \| Ding, \| Qiang Luo, \| and \| Guorui \| Zhou. \|
	\| --- \| --- \| --- \| --- \| --- \| --- \| --- \| ----------- \| --- \| ----- \| ---------- \| --- \| ------ \| ----- \|
	bytheirrespectivelicenses;forreproducibility,the 2025. Onerec: Unifyingretrieveandrankwithgen-
	officialdatasetsourcesarecitedinthecorrespond- erativerecommenderanditerativepreferencealign-
	ment. Preprint,arXiv:2502.18965.
	ingsectionsofthispaper,andwedonotredistribute
	anyrawuser-generatedcontentsuchasreviewtext AbhimanyuDubey,AbhinavJauhri,AbhinavPandey,
	\| aspartofthiswork. \| \| Thisresearchdoesnotattempt \| \| \| \| \| \| \| \| \| \| \| \|
	\| ----------------- \| --- \| -------------------------- \| --- \| --- \| --- \| --- \| --- \| --- \| --- \| --- \| --- \| --- \| --- \|
	AbhishekKadian,AhmadAl-Dahle,AieshaLetman,
	to identify or infer the identity of any individual Akhil Mathur, Alan Schelten, Amy Yang, Angela
	\| \| \| \| \| \| \| \| Fan,and1others.2024. \| \| \| Thellama3herdofmodels. \| \| \| \|
	\| --- \| --- \| --- \| --- \| --- \| --- \| --- \| -------------------- \| --- \| --- \| ---------------------- \| --- \| --- \| --- \|
	user,andpriortotraininganddownstreamprofile
	arXivpreprintarXiv:2407.21783.
	\| generation, \| we \| apply text \| filtering \| \| procedures \| to \| \| \| \| \| \| \| \|
	\| ----------- \| --- \| ---------- \| --------- \| --- \| ---------- \| --- \| --- \| --- \| --- \| --- \| --- \| --- \| --- \|
	remove direct personal identifiers where applica- YiFang,WenjieWang,YangZhang,FengbinZhu,Qi-
	ble;themodelsaretrainedsolelyonpubliclyavail- fanWang,FuliFeng,andXiangnanHe.2025. Rea-
	son4rec:Largelanguagemodelsforrecommendation
	abledata,andnoadditionalprivateorproprietary
	\| \| \| \| \| \| \| \| withdeliberativeuserpreferencealignment. \| \| \| \| \| Preprint, \| \|
	\| ---------------- \| --- \| -------------- \| --- \| ------------ \| --- \| --- \| ---------------------------------------- \| --- \| --- \| --- \| --- \| --------- \| --- \|
	\| user information \| \| is introduced. \| \| Furthermore, \| \| we \| \| \| \| \| \| \| \|
	arXiv:2502.02061.
	\| do not release \| generated \| \| user \| profiles \| or \| any per- \| \| \| \| \| \| \| \|
	\| -------------- \| --------- \| --- \| ---- \| -------- \| --- \| -------- \| --- \| --- \| --- \| --- \| --- \| --- \| --- \|
	MinghuaHe,YueChen,FangkaiYang,PuZhao,Wenjie
	useroutputsproducedduringthemodelingprocess,
	Yin,YuKang,QingweiLin,SaravanRajmohan,and
	andallreportedresultsarepresentedinaggregate
	\| \| \| \| \| \| \| \| DongmeiZhang.2025. \| \| \| Execoder:Empoweringlarge \| \| \| \|
	\| -------- \| ---------- \| -------- \| --- \| ----- \| ---- \| ------- \| ------------------ \| --- \| --- \| ------------------------ \| --- \| --- \| --- \|
	\| form for \| evaluation \| purposes \| \| only. \| From \| a secu- \| \| \| \| \| \| \| \|
	languagemodelswithexecutabilityrepresentationfor
	rityperspective,theassociatedcoderepositoryhas codetranslation. Preprint,arXiv:2501.18460.
	beenscannedforpotentialsecrets,andanyinternal
	MinjieHong,ZirunGuo,YanXia,ZehanWang,Ziang
	endpointsorconfidentialconfigurationinformation
	\| \| \| \| \| \| \| \| Zhang, \| Tao \| Jin, and \| Zhou Zhao. \| 2025a. \| Apo: \| En- \|
	\| --- \| --- \| --- \| --- \| --- \| --- \| --- \| ------ \| --- \| -------- \| ---------- \| ------ \| ---- \| --- \|
	havebeenremovedpriortorelease.
	hancingreasoningabilityofmllmsviaasymmetric
	\| \| \| \| \| \| \| \| policyoptimization. \| \| \| Preprint,arXiv:2506.21655. \| \| \| \|
	\| --- \| --- \| --- \| --- \| --- \| --- \| --- \| ------------------- \| --- \| --- \| -------------------------- \| --- \| --- \| --- \|

	Minjie Hong, Yan Xia, Zehan Wang, Jieming Zhu, JoarSkalse,NikolausH.R.Howe,DmitriiKrashenin-
	Ye Wang, Sihang Cai, Xiaoda Yang, Quanyu Dai, nikov, and David Krueger. 2025. Defining
	ZhenhuaDong,ZhimengZhang,and1others.2025b. and characterizing reward hacking. Preprint,
	\| Eager-llm: \| Enhancinglargelanguagemodelsasrec- \| \| \| \| \| arXiv:2209.13085. \| \| \| \| \| \|
	\| ---------- \| ---------------------------------- \| --------- \| ----------------- \| --- \| --- \| ----------------- \| --- \| --- \| --- \| --- \| --- \|
	\| ommenders \| through \| exogenous \| behavior-semantic \| \| \| \| \| \| \| \| \|
	integration. InProceedingsoftheACMonWebCon- Harald Steck. 2019. Embarrassingly shallow autoen-
	ference2025,pages2754–2762. codersforsparsedata. InProceedingsofTheWorld
	WideWebConference(WWW),pages3251–3257.
	MinjieHong,ZetongZhou,ZirunGuo,ZiangZhang,
	\| Ruofan \| Hu, Weinan \| Gan, \| Jieming \| Zhu, and \| Zhou \| \| \| \| \| \| \|
	\| ------ \| ---------- \| ---- \| ------- \| -------- \| ---- \| --- \| --- \| --- \| --- \| --- \| --- \|
	ChaoSun,YaoboLiang,YamingYang,ShilinXu,Tian-
	\| Zhao.2025c. \| Generativereasoningrecommendation \| \| \| \| \| \| \| \| \| \| \|
	\| ----------- \| --------------------------------- \| --- \| --- \| --- \| --- \| ---------------------------- \| -------- \| ---- \| ------ \| -------- \| ------------- \|
	\| \| \| \| \| \| \| mengYang,andYunhaiTong.2024. \| \| \| \| \| Rlrf4rec: Re- \|
	\| viallms. \| Preprint,arXiv:2510.20815. \| \| \| \| \| \| \| \| \| \| \|
	\| \| \| \| \| \| \| inforcement \| learning \| from \| recsys \| feedback \| for en- \|
	JihwanJeong,YinlamChow,GuyTennenholtz,Chih- hancedrecommendationreranking. arXivpreprint
	\| Wei Hsu, \| Azamat \| Tulepbergenov, \| \| Mohammad \| \| arXiv:2410.05939. \| \| \| \| \| \|
	\| ------------------------------------ \| ------------ \| --------------- \| ---------- \| -------- \| --------- \| ----------------- \| --- \| --------------------- \| --- \| --- \| --------- \|
	\| Ghavamzadeh, \| and \| Craig \| Boutilier. \| 2023. \| Fac- \| \| \| \| \| \| \|
	\| \| \| \| \| \| \| QwenTeam.2025. \| \| Qwen3technicalreport. \| \| \| Preprint, \|
	\| tual and \| personalized \| recommendations \| \| using \| lan- \| \| \| \| \| \| \|
	\| guagemodelsandreinforcementlearning. \| \| \| \| \| Preprint, \| arXiv:2505.09388. \| \| \| \| \| \|
	arXiv:2310.06176.
	\| \| \| \| \| \| \| Hugo Touvron, \| Louis \| Martin, \| Kevin \| Stone, \| Peter Al- \|
	\| --- \| --- \| --- \| --- \| --- \| --- \| ------------- \| ----- \| ------- \| ----- \| ------ \| --------- \|
	Yitong Ji, Aixin Sun, Jie Zhang, and Chenliang Li. bert, Amjad Almahairi, Yasmine Babaei, Nikolay
	\| 2023. A \| critical \| study \| on data leakage \| in \| recom- \| \| \| \| \| \| \|
	\| ------- \| -------- \| ----- \| --------------- \| --- \| ------ \| --- \| --- \| --- \| --- \| --- \| --- \|
	Bashlykov,SoumyaBatra,PrajjwalBhargava,Shruti
	\| mendersystemofflineevaluation. \| \| \| \| ACMTrans.Inf. \| \| \| \| \| \| \| \|
	\| ------------------------------ \| --- \| --- \| --- \| ------------- \| --- \| ------------------------ \| --- \| --- \| ------- \| --- \| --------- \|
	\| \| \| \| \| \| \| Bhosale,and1others.2023. \| \| \| Llama2: \| \| Openfoun- \|
	Syst.,41(3):75:1–75:27.
	\| \| \| \| \| \| \| dation and \| fine-tuned \| chat \| models. \| arXiv \| preprint \|
	\| --- \| --- \| --- \| --- \| --- \| --- \| ---------- \| ---------- \| ---- \| ------- \| ----- \| -------- \|
	arXiv:2307.09288.
	\| JiachengLin,TianWang,andKunQian.2025. \| \| \| \| \| Rec-r1: \| \| \| \| \| \| \|
	\| ------------------------------------- \| --- \| --- \| --- \| --- \| ------- \| --- \| --- \| --- \| --- \| --- \| --- \|
	Bridginggenerativelargelanguagemodelsanduser-
	\| \| \| \| \| \| \| LeandrovonWerra, \| \| YounesBelkada, \| \| LewisTunstall, \| \|
	\| --- \| --- \| --- \| --- \| --- \| --- \| ---------------- \| --- \| -------------- \| --- \| -------------- \| --- \|
	centricrecommendationsystemsviareinforcement
	EdwardBeeching,TristanThrush,NathanLambert,
	\| learning. \| Preprint,arXiv:2503.24289. \| \| \| \| \| \| \| \| \| \| \|
	\| --------- \| -------------------------- \| --- \| --- \| --- \| --- \| -------- \| ------ \| ----------- \| ------------- \| --- \| ------------ \|
	\| \| \| \| \| \| \| Shengyi \| Huang, \| Kashif \| Rasul, \| and \| Quentin Gal- \|
	\| \| \| \| \| \| \| louédec. \| TRL: \| Transformer \| Reinforcement \| \| Learn- \|
	AiweiLiu,MinghuaHe,ShaoxunZeng,SijunZhang,
	\| LinhaoZhang,ChuhanWu,WeiJia,YuanLiu,Xiao \| \| \| \| \| \| ing. \| \| \| \| \| \|
	\| ---------------------------------------- \| --- \| --- \| ----------------------- \| --- \| --- \| ---- \| --- \| --- \| --- \| --- \| --- \|
	\| Zhou,andJieZhou.2025. \| \| \| Wedlm:Reconcilingdiffu- \| \| \| \| \| \| \| \| \|
	sionlanguagemodelswithstandardcausalattention LuWang,DiZhang,FangkaiYang,PuZhao,Jianfeng
	forfastinference. Preprint,arXiv:2512.22737. Liu,YuefengZhan,HaoSun,QingweiLin,Weiwei
	\| \| \| \| \| \| \| Deng, Dongmei \| \| Zhang, \| Feng \| Sun, and \| Qi Zhang. \|
	\| --- \| --- \| --- \| --- \| --- \| --- \| ------------- \| --- \| ------ \| ---- \| -------- \| --------- \|
	HongtaoLiu,FangzhaoWu,WenjunWang,Xianchen
	\| \| \| \| \| \| \| 2025. Lettingo: \| \| Exploreuserprofilegenerationfor \| \| \| \|
	\| --- \| --- \| --- \| --- \| --- \| --- \| --------------- \| --- \| ------------------------------- \| --- \| --- \| --- \|
	Wang,PengfeiJiao,ChuhanWu,andXingXie.2019.
	\| \| \| \| \| \| \| recommendationsystem. \| \| \| InProceedingsofthe31st \| \| \|
	\| --- \| --- \| --- \| --- \| --- \| --- \| --------------------- \| --- \| --- \| ---------------------- \| --- \| --- \|
	Nrpa: Neuralrecommendationwithpersonalizedat- ACMSIGKDDConferenceonKnowledgeDiscovery
	tention. In Proceedings of the 42nd International andDataMiningV.2,KDD’25,NewYork,NY,USA.
	ACMSIGIRConferenceonResearchandDevelop- AssociationforComputingMachinery.
	mentinInformationRetrieval,pages1233–1236.
	YeWang,JiahaoXun,MinjieHong,JiemingZhu,Tao
	\| Wensheng \| Lu, Jianxun \| Lian, \| Wei Zhang, \| Guanghua \| \| \| \| \| \| \| \|
	\| -------- \| ----------- \| ----- \| ---------- \| -------- \| --- \| --- \| --- \| --- \| --- \| --- \| --- \|
	Jin,WangLin,HaoyuanLi,LinjunLi,YanXia,Zhou
	Li,MingyangZhou,HaoLiao,andXingXie.2024.
	\| \| \| \| \| \| \| Zhao,and1others.2024. \| \| \| Eager: \| Two-streamgen- \| \|
	\| --- \| --- \| --- \| --- \| --- \| --- \| --------------------- \| --- \| --- \| ------ \| -------------- \| --- \|
	Aligninglargelanguagemodelsforcontrollablerec-
	arXivpreprintarXiv:2403.05063. erativerecommenderwithbehavior-semanticcollab-
	ommendations.
	\| \| \| \| \| \| \| oration. \| InProceedingsofthe30thACMSIGKDD \| \| \| \| \|
	\| --- \| --- \| --- \| --- \| --- \| --- \| -------- \| ------------------------------- \| --- \| --- \| --- \| --- \|
	RadaMihalceaandPaulTarau.2004. Textrank: Bring- ConferenceonKnowledgeDiscoveryandDataMin-
	\| \| \| InEMNLP. \| \| \| \| ing,pages3245–3254. \| \| \| \| \| \|
	\| --- \| --- \| -------- \| --- \| --- \| --- \| ------------------- \| --- \| --- \| --- \| --- \| --- \|
	ingorderintotext.
	\| Zhaopeng \| Qiu, Xian \| Wu, \| Jingyue Gao, \| and \| Wei Fan. \| \| \| \| \| \| \|
	\| -------- \| --------- \| --- \| ------------ \| --- \| -------- \| -------- \| --------------- \| --- \| -------- \| -------- \| --- \|
	\| \| \| \| \| \| \| Triyanna \| Widiyaningtyas, \| \| Indriana \| Hidayah, \| and \|
	2021. U-BERT:pre-traininguserrepresentationsfor
	\| \| \| \| \| \| \| TeguhBAdji.2021. \| \| Userprofilecorrelation-based \| \| \| \|
	\| --- \| --- \| --- \| --- \| --- \| --- \| ---------------- \| --- \| ---------------------------- \| --- \| --- \| --- \|
	improved recommendation. In AAAI, pages 4320– similarity(upcsim)algorithminmovierecommenda-
	\| 4327.AAAIPress. \| \| \| \| \| \| tionsystem. \| JournalofBigData,8(1):52. \| \| \| \| \|
	\| --------------------------------- \| --- \| --- \| --- \| -------------- \| --- \| ----------- \| ----------------------------------- \| --- \| --- \| --- \| --- \|
	\| NilsReimersandIrynaGurevych.2019. \| \| \| \| Sentence-bert: \| \| \| \| \| \| \| \|
	\| \| \| \| \| \| \| J.Wu.2023. \| Computationalunderstandingofuserin- \| \| \| \| \|
	Sentenceembeddingsusingsiamesebert-networks.
	terfaces.
	InProceedingsofthe2019ConferenceonEmpirical
	MethodsinNaturalLanguageProcessing(EMNLP),
	pages3982–3992. Yunjia Xi, Weiwen Liu, Jianghao Lin, Xiaoling Cai,
	\| \| \| \| \| \| \| Hong Zhu, \| Jieming \| Zhu, \| Bo Chen, \| Ruiming \| Tang, \|
	\| --- \| --- \| --- \| --- \| --- \| --- \| --------- \| ------- \| ---- \| -------- \| ------- \| ----- \|
	Xubin Ren, Wei Wei, Lianghao Xia, Lixin Su, Suqi WeinanZhang,andYongYu.2024. Towardsopen-
	Cheng,JunfengWang,DaweiYin,andChaoHuang. world recommendation with knowledge augmenta-
	2024. Representationlearningwithlargelanguage tionfromlargelanguagemodels. InProceedingsof
	modelsforrecommendation. InProceedingsofthe the18thACMConferenceonRecommenderSystems,
	\| ACMonWebConference2024,pages3464–3475. \| \| \| \| \| \| pages12–22. \| \| \| \| \| \|
	\| -------------------------------------- \| --- \| --- \| --- \| --- \| --- \| ----------- \| --- \| --- \| --- \| --- \| --- \|

	A ExperimentSetup
	FanYang,ZhengChen,ZiyanJiang,EunahCho,Xiao-
	\| jiangHuang,andYanbinLu.2023. \| \| Palr:Personaliza- \| \| \| \| \| \| \| \|
	\| ------------------------------- \| --- \| ----------------- \| ---------------- \| --- \| --- \| --- \| --- \| --- \| --- \|
	\| tionawarellmsforrecommendation. \| \| arXivpreprint \| A.1 DataCuration \| \| \| \| \| \| \|
	arXiv:2305.07622.
	Jiarui Zhang. 2024. Guided profile generation im- Table7: Statisticaldetailsoftheevaluationdatasets.
	\| proves personalization \| with llms. \| arXiv preprint \| \| \| \| \| \| \| \|
	\| ---------------------- \| ---------- \| -------------- \| ---------- \| ------ \| ----------- \| --- \| ------------ \| ------ \| ------ \|
	\| arXiv:2409.13093. \| \| \| Dataset \| #Train \| #Valid \| \| #Test #User \| \| #Item \|
	\| \| \| \| Music \| 43,071 \| 3,271 \| \| 1,296 \| 4,183 \| 2,660 \|
	\| \| \| \| Book \| 71,972 \| 6,144 \| \| 5,541 13,863 \| \| 13,515 \|
	\| \| \| \| Yelp \| 51,497 \| 4,757 \| \| 4,328 \| 8,453 \| 13,426 \|
	\| \| \| \| We conduct \| \| experiments \| \| on three \| widely \| used \|
	real-worlddatasets:
	\| \| \| \| • AmazonMusic(Music): \| \| \| \| Thisreferstothe“Dig- \| \| \|
	\| --- \| --- \| --- \| --------------------- \| --- \| ------ \| ------ \| -------------------- \| --- \| ------ \|
	\| \| \| \| ital Music” \| \| subset \| of the \| well-known \| \| Amazon \|
	Productdataset3,whichrecordsrichuserreviews,
	ratings,andtextualinformationaboutitems,such
	\| \| \| \| as titles, \| across \| a \| broad \| range \| of product \| cate- \|
	\| --- \| --- \| --- \| ---------- \| ------ \| --- \| ----- \| ----- \| ---------- \| ----- \|
	gories,ontheAmazonplatform.2
	\| \| \| \| • AmazonBook(Book): \| \| \| Thisreferstothe“Book” \| \| \| \|
	\| --- \| --- \| --- \| ------------------- \| --- \| --- \| --------------------- \| --- \| --- \| --- \|
	subsetoftheAmazonProductdataset.
	dataset4,
	\| \| \| \| • Yelp: \| This \| refers \| to the \| Yelp \| Open \| \|
	\| --- \| --- \| --- \| ------- \| -------- \| -------------- \| -------- \| ------- \| ------------- \| --------- \|
	\| \| \| \| which \| includes \| user \| reviews, \| ratings \| \| for busi- \|
	\| \| \| \| nesses \| such \| as restaurants \| \| and \| retail shops, \| as \|
	wellastextualinformationaboutthebusinesses.
	Itiswidelyusedinrecommendationtasks(Qiu
	etal.,2021).
	WeusetheentireMusicdatasetforexperiments,
	\| \| \| \| while for \| the Book \| \| and Yelp \| datasets, \| we \| utilize \|
	\| --- \| --- \| --- \| ------------------------------- \| -------- \| --- \| -------- \| --------- \| ---------- \| ------- \|
	\| \| \| \| onlyasubsetduetotheirlargesize. \| \| \| \| \| FortheBook \| \|
	dataset,weusedatafromthelasttwomonths,and
	fortheYelpdataset,weusedatafromthelastsix
	months. Foreachdataset,wesplititintotraining,
	\| \| \| \| validation, \| and \| test \| sets based \| on \| the timestamps \| \|
	\| --- \| --- \| --- \| ---------------- \| --- \| -------- \| ---------- \| ---- \| -------------- \| --- \|
	\| \| \| \| of interactions, \| \| ensuring \| that \| test \| interactions \| oc- \|
	curafteralltrainingandvalidationinteractionsto
	preventinformationleakage(Jietal.,2023).
	\| \| \| \| Regarding \| \| data \| filtering, \| following \| \| prior \|
	\| --- \| --- \| --- \| --------- \| --- \| ----------- \| ---------- \| --------- \| -------- \| ----- \|
	\| \| \| \| work (Liu \| et \| al., 2019), \| we \| adopt \| a 5-core \| set- \|
	tingtofilterthedataandexcludecold-startusers
	\| \| \| \| and items—those \| \| not \| appearing \| \| in the \| training \|
	\| --- \| --- \| --- \| --------------- \| --- \| ---------- \| --------- \| ---- \| ------ \| -------- \|
	\| \| \| \| set—from \| the \| validation \| and \| test \| sets. \| The sta- \|
	tisticaldetailsoftheprocesseddatasetareprovided
	inTable7.
	3https://cseweb.ucsd.edu/~jmcauley/datasets/
	amazon/links.html.
	4https://business.yelp.com/data/resources/
	open-dataset/.

	\| A.1.1 \| ImplementationDetails \| \| \| \| \| \| RLMRecPrompt \| \| \| \|
	\| ----- \| --------------------- \| --- \| --- \| --- \| --- \| --- \| ------------ \| --- \| --- \| --- \|
	Inourexperiments,weprimarilyemployQwen3- Role: Businessrecommendationassistant
	Task: Determinebusinesstypesauserislikely
	\| 8B(Team,2025)andLLaMA38BInstruct \| \| \| \| \| \| (Dubey \| \| \| \| \|
	\| ------------------------------------------ \| --- \| ----- \| --- \| --- \| ------- \| ---------- \| ------------ \| --- \| --- \| --- \|
	\| etal.,2024;Touvronetal.,2023)asboththerec- \| \| \| \| \| \| \| toenjoy \| \| \| \|
	\| ommendation \| \| model \| and \| the \| profile \| generation \| InputFormat: \| \| \| \|
	•Title: Businessname
	model. Thetrainingprocessisimplementedusing
	theTRL(vonWerraetal.). Keyhyperparameters, •Categories: Businesscategories
	suchasbatchsizeandlearningrate,aredetermined •Sentiment: Usersentimenttowardbusiness
	throughgridsearchtoachieveoptimalperformance.
	OutputRequirements:
	Moredetailscanbefoundinourcode.
	1. JSONformatonly
	2. Structure:
	A.2 BaselinePrompts
	{
	\| \| \| \| \| \| \| \| "summarization": \| "Types \| of businesses \| user \|
	\| --- \| --- \| --- \| --- \| --- \| --- \| --- \| ---------------- \| ------ \| ------------- \| ---- \|
	KARPrompt
	likelyenjoys"(≤100words),
	Task: Analyzeuserpreferencesbasedonbusi-
	\| \| \| \| \| \| \| \| "reasoning": \| "Brief explanation \| \| for summa- \|
	\| --- \| --- \| --- \| --- \| --- \| --- \| --- \| ------------ \| ------------------ \| --- \| ---------- \|
	nessreviewinghistory
	rization"(nowordlimit)
	Input: {user_history}-User’sbusinessreview-
	}
	inghistorywithsentimentsovertime
	3. NoadditionaltextoutsideJSON
	Instructions:
	\| 1. \| Analyze \| the \| user’s \| preferences \| considering \| \| \| \| \| \|
	\| --- \| ------- \| --- \| ------ \| ----------- \| ----------- \| --- \| ------ \| ----------- \| --- \| ------ \|
	\| \| \| \| \| \| \| \| Input: \| INTERACTION \| \| ITEMS: \|
	businessnamesandcategories
	{user_history}
	\| 2. \| Take \| into account \| \| sentiment \| patterns \| over \| \| \| \| \|
	\| --- \| ---- \| ------------ \| --- \| --------- \| -------- \| ---- \| --- \| --- \| --- \| --- \|
	time
	\| 3. \| Provideclearexplanationsbasedonreview- \| \| \| \| \| \| LGPrompt \| \| \| \|
	\| --- \| -------------------------------------- \| --- \| --- \| --- \| --- \| --- \| -------- \| --- \| --- \| --- \|
	inghistorydetails
	Youwillserveasanassistanttohelpmegener-
	\| 4. \| Consider \| other \| pertinent \| \| factors \| that may \| \| \| \| \|
	\| --- \| -------- \| ----- \| --------- \| --- \| ------- \| -------- \| --- \| --- \| --- \| --- \|
	ateauserprofilebasedonthisuser’ssentiments
	influencepreferences
	historytobetterunderstandthisusers’interest
	andthuspredicthis/hersentimentaboutatarget
	\| PALRPrompt \| \| \| \| \| \| \| item. Iwillprovideyouwithsomebehaviorhis- \| \| \| \|
	\| ---------- \| --- \| --- \| --- \| --- \| --- \| --- \| ----------------------------------------- \| --- \| --------------- \| --- \|
	\| \| \| \| \| \| \| \| toryoftheuserinthisformat: \| \| [itemattributes \| \|
	Task: Summarizeuserpreferencesusingkey-
	\| \| \| \| \| \| \| \| andsentiment]. \| Theuserprofileyougenerate \| \| \|
	\| --- \| --- \| --- \| --- \| --- \| --- \| --- \| -------------- \| ------------------------- \| --- \| --- \|
	words.
	shouldcontainasmuchusefulcontentaspossi-
	\| Input: \| \| {user_history} \| \| - historical \| businesses \| \| \| \| \| \|
	\| ------ \| --- \| -------------- \| --- \| ------------ \| ---------- \| --- \| --- \| --- \| --- \| --- \|
	bletohelppredicttheuser’ssentimenttowards
	withusersentiments.
	anewbusiness.
	\| Output \| \| Format: \| An \| itemized \| list ranked \| by \| \| \| \| \|
	\| ------ \| --- \| ------- \| --- \| -------- \| ----------- \| --- \| --- \| --- \| --- \| --- \|
	USERHISTORY:{user_history}.
	importance.
	PROFILEYOUGENERATE:
	Template:
	\| \| • KEY_WORD_1: \| \| \| \| \| "HIS- \| \| \| \| \|
	\| --- \| ------------- \| --- \| --- \| --- \| --- \| ----- \| --- \| --- \| --- \| --- \|
	R4RecPrompt(Reasoner)
	\| \| TORY_BUSINESS_1", \| \| \| \| \| "HIS- \| \| \| \| \|
	\| --- \| ----------------- \| --- \| --- \| --- \| --- \| ----- \| --- \| --- \| --- \| --- \|
	UserReviewHistory
	TORY_BUSINESS_2"
	⟨H u organizedasbelow⟩
	\| \| • KEY_WORD_2: \| \| \| \| \| "HIS- \| 1. TitleofItem1 \| \| \| \|
	\| ------------- \| ---------------- \| --- \| --- \| --- \| --- \| ----- \| ---------------- \| ----------------------- \| ----------- \| --- \|
	\| \| TORY_BUSINESS_3" \| \| \| \| \| \| PositiveAspects: \| [Aspect1],[Aspect2],... \| \| \|
	\| \| \| \| \| \| \| \| NegativeAspects: \| [Aspect1],[Aspect2],... \| \| \|
	\| Instructions: \| \| \| \| \| \| \| User Preference \| Elements: \| [Preference \| 1], \|
	[Preference2],...
	\| \| 1. Extract \| key \| preference \| \| indicators \| from \| \| \| \| \|
	\| --- \| ---------- \| --- \| ---------- \| --- \| ---------- \| ---- \| --- \| --- \| --- \| --- \|
	2. TitleofItem2
	userinteractionhistory.
	\| \| \| \| \| \| \| \| PositiveAspects: \| [Aspect1],[Aspect2],... \| \| \|
	\| --- \| --- \| --- \| --- \| --- \| --- \| --- \| ---------------- \| ----------------------- \| --- \| --- \|
	2. Rankkeywordsbyimportance. NegativeAspects: [Aspect1],[Aspect2],...
	\| \| \| \| \| \| \| \| User Preference \| Elements: \| [Preference \| 1], \|
	\| --- \| --- \| --- \| --- \| --- \| --- \| --- \| --------------- \| --------- \| ----------- \| --- \|

	\| [Preference2],... \| \| \| \| \| \| \| cantlyworsethanDUETacrossalldatasets. \| \| \| \| \| This \|
	\| ----------------- \| --- \| --- \| --- \| --- \| --- \| --- \| ------------------------------------- \| --- \| --- \| --- \| --- \| ---- \|
	indicatesthatcoherentsummariesaloneareinsuf-
	...
	\| \| \| \| \| \| \| \| ficient, \| and that \| the gains \| of DUET \| \| arise from \|
	\| --- \| --- \| --- \| --- \| --- \| --- \| --- \| -------- \| -------- \| --------- \| ------- \| --- \| ---------- \|
	ItemReviewHistorybyOtherUsers
	⟨H organizedinthesameformatasabove⟩ learnedsemanticabstractionratherthanextractive
	i
	compression.
	\| Task: \| Analyze \| whether \| \| the user \| will \| like the \| \| \| \| \| \| \|
	\| --------- \| ------- \| ------- \| --- \| ---------- \| ----------- \| -------- \| --- \| --- \| --- \| --- \| --- \| --- \|
	\| new Music \| \| i based \| on \| the user’s \| preferences \| \| \| \| \| \| \| \|
	B.2 RankingunderHardNegativeSampling
	\| andtheitem’sfeatures. \| \| \| \| Provideyourrationale \| \| \| \| \| \| \| \| \|
	\| --------------------- \| --- \| --- \| --- \| -------------------- \| --- \| --- \| --- \| --- \| --- \| --- \| --- \| --- \|
	Toconstructamorechallengingrankingscenario,
	inoneconciseparagraph.
	wereplacerandomnegativeswithhardnegatives
	\| \| \| \| \| \| \| \| generatedbyacollaborativefilteringmodel. \| \| \| \| \| Specif- \|
	\| --- \| --- \| --- \| --- \| --- \| --- \| --- \| ---------------------------------------- \| --- \| --- \| --- \| --- \| ------- \|
	A.3 ExampleofCue
	ically,foreachuser,weretrievehigh-scoringitems
	Thefollowingexamplesillustratehowrawsignals
	\| \| \| \| \| \| \| \| from an \| EASE(Steck, \| 2019) \| model \| that \| the user \|
	\| --------------------- \| --- \| --- \| ---------------- \| --- \| --- \| ---- \| ------- \| ----------- \| ----- \| ----- \| ---- \| -------- \|
	\| aredistilledintocues. \| \| \| Asshownonthetop, \| \| \| user \| \| \| \| \| \| \|
	hasnotinteractedwith,andcombinethemwiththe
	cuesemphasizehistoricalpreferences,whileitem
	ground-truthitemtoformthecandidateset.
	cues highlight metadata and user-group patterns. Table 9 reports the results. Compared to ran-
	\| Together, \| they \| provide \| minimal \| \| but informative \| \| \| \| \| \| \| \|
	\| --------- \| ---- \| ------- \| ------- \| --- \| --------------- \| --- \| --- \| --- \| --- \| --- \| --- \| --- \|
	domsampling,performancedecreasesforallmeth-
	hypothesesforprofileexploration.
	\| \| \| \| \| \| \| \| ods due \| to increased \| difficulty, \| while \| DUET \| con- \|
	\| --- \| --- \| --- \| --- \| --- \| --- \| --- \| --------- \| ------------ \| ----------- \| ---------------- \| ---- \| ------ \|
	\| \| \| \| \| \| \| \| sistently \| maintains \| the \| best performance \| \| across \|
	ExamplesofUserCues
	\| \| \| \| \| \| \| \| datasets. \| Thisindicatesthattheimprovementsare \| \| \| \| \|
	\| ------- \| ----- \| ------ \| ------ \| --- \| ------- \| -------- \| --------- \| ----------------------------------- \| --- \| --- \| --- \| --- \|
	\| “enjoys \| retro \| puzzle \| games” \| — \| derived \| from re- \| \| \| \| \| \| \|
	robustandnotlimitedtotrivialrankingscenarios.
	peatedengagementwithclassictitles.
	“prefersconciseproductreviews”—inferredfrom B.3 RobustnessunderPreferenceDiversity
	apatternofshort,directcomments.
	WefurtheranalyzetherobustnessofDUETunder
	“tendstogivehighratingsbutrarelycomments”—
	\| \| \| \| \| \| \| \| varyinglevelsofuserpreferencediversity. \| \| \| \| \| Weuse \|
	\| --- \| --- \| --- \| --- \| --- \| --- \| --- \| --------------------------------------- \| --- \| --- \| --- \| --- \| ----- \|
	highlightingconsistencybutlimitedfeedback.
	thevarianceofhistoricalratingsasaproxyforpref-
	\| \| \| \| \| \| \| \| erencestability: \| \| lowvarianceindicatesconsistent \| \| \| \|
	\| --- \| --- \| --- \| --- \| --- \| --- \| --- \| ---------------- \| --- \| ------------------------------ \| --- \| --- \| --- \|
	ExamplesofItemCues
	\| \| \| \| \| \| \| \| preferences, \| while \| high \| variance \| corresponds \| to \|
	\| --- \| --- \| --- \| --- \| --- \| --- \| --- \| ------------ \| ----- \| ---- \| -------- \| ----------- \| --- \|
	“lightweighttrail-runningshoes”—derivedfrom
	diverseorpotentiallyconflictingsignals.
	productmetadata.
	Wepartitionusersintothreegroupsbasedonper-
	\| “popular \| among \| budget-conscious \| \| \| users” \| — in- \| \| \| \| \| \| \|
	\| -------- \| ----- \| ---------------- \| --- \| --- \| ------ \| ----- \| --- \| --- \| --- \| --- \| --- \| --- \|
	centilethresholds(bottom33%,middle33%,top
	ferredfrompurchasepatterns.
	33%ofratingvariance)andevaluateperformance
	\| “stylized \| with \| retro \| aesthetics” \| \| — extracted \| from \| \| \| \| \| \| \|
	\| --------- \| ---- \| ----- \| ----------- \| --- \| ----------- \| ---- \| --- \| --- \| --- \| --- \| --- \| --- \|
	withineachgroup.
	itemdescriptions.
	\| \| \| \| \| \| \| \| As shown \| in \| Table 10, \| performance \| \| degrades \|
	\| --- \| --- \| --- \| --- \| --- \| --- \| --- \| -------- \| --- \| --------- \| ----------- \| --- \| -------- \|
	smoothlyaspreferencediversityincreasesacross
	\| \| \| \| \| \| \| \| all datasets. \| Importantly, \| \| the degradation \| \| is grad- \|
	\| --- \| --- \| --- \| --- \| --- \| --- \| --- \| ------------- \| ------------ \| --- \| --------------- \| --- \| -------- \|
	B AdditionalExperiments
	ualratherthancatastrophic,indicatingthatDUET
	B.1 Non-LLMBaselineviaExtractive
	remainsstableunderheterogeneousornoisyinter-
	Summarization
	actionhistories.
	\| To examine \| \| whether \| the \| gains \| of DUET \| stem \| \| \| \| \| \| \|
	\| ---------- \| --- \| ------- \| --- \| ----- \| ------- \| ---- \| --- \| --- \| --- \| --- \| --- \| --- \|
	C TheUseofLargeLanguageModels
	\| from improved \| \| semantic \| \| representations \| \| rather \| \| \| \| \| \| \|
	\| ------------- \| --- \| ---------------- \| --- \| --------------- \| --------- \| ------ \| --- \| --- \| --- \| --- \| --- \| --- \|
	\| than generic \| \| text generation, \| \| we \| introduce \| a non- \| \| \| \| \| \| \|
	WeusedaLargeLanguageModel(LLM)onlyas
	LLMbaselinebasedonextractivesummarization. a writing assistant to polish the language of the
	\| Specifically, \| \| we apply \| TextRank \| \| (Mihalcea \| and \| \| \| \| \| \| \|
	\| ------------- \| --- \| -------- \| -------- \| --- \| --------- \| --- \| ---------- \| ------ \| ------- \| ----------- \| --- \| --------- \|
	\| \| \| \| \| \| \| \| manuscript \| (e.g., \| grammar \| refinement, \| \| style ad- \|
	Tarau,2004)toselectsalientsentencesfromuser
	\| \| \| \| \| \| \| \| justment,andclarityimprovement). \| \| \| \| Theresearch \| \|
	\| --- \| --- \| --- \| --- \| --- \| --- \| --- \| -------------------------------- \| --- \| --- \| --- \| ----------- \| --- \|
	historiesandconstructuserprofileswithoutusing ideas,methodologydesign,experiments,andanal-
	anygenerativemodel. ysiswereentirelyconceived,implemented,andval-
	\| The \| extracted \| summaries \| \| are \| then fed \| into the \| \| \| \| \| \| \|
	\| --- \| --------- \| --------- \| --- \| --- \| -------- \| -------- \| --- \| --- \| --- \| --- \| --- \| --- \|
	idatedbytheauthorswithoutrelianceontheLLM.
	samedownstreampredictorforratingestimation.
	\| \| \| \| \| \| \| \| The LLM \| did not \| contribute \| to research \| \| ideation, \|
	\| --- \| --- \| --- \| --- \| --- \| --- \| --- \| ------- \| ------- \| ---------- \| ----------- \| --- \| --------- \|
	Thisbaselineisolatestheeffectofreadablesumma- experimentaldesign,orresultinterpretation.
	rizationfromrepresentationlearning.
	Table8showsthatTextRankimprovesoversim-
	\| ple history \| truncation \| \| (10H), \| but \| remains \| signifi- \| \| \| \| \| \| \|
	\| ----------- \| ---------- \| --- \| ------ \| --- \| ------- \| -------- \| --- \| --- \| --- \| --- \| --- \| --- \|

	\| Method \| \| Yelp \| \| \| AmazonMusic \| \| \| AmazonBooks \| \|
	\| ------ \| -------- \| ---- \| ----- \| --- \| ----------- \| --------- \| --- \| ----------- \| --------- \|
	\| \| MAE RMSE \| Acc \| F1(%) \| MAE \| RMSE \| Acc F1(%) \| MAE \| RMSE \| Acc F1(%) \|
	\| \| \| (%) \| \| \| \| (%) \| \| \| (%) \|
	10H 0.8283 1.3893 48.53 41.15 0.7322 1.1430 57.18 52.48 0.8741 1.4760 51.83 50.35
	TextRank 0.8104 1.1323 50.51 30.23 0.5914 0.8223 61.19 29.68 0.6328 0.8854 59.59 29.48
	Ours 0.5126 0.9485 61.23 55.18 0.3937 0.7564 67.96 63.89 0.4612 0.9089 64.38 59.27
	Table8: Comparisonwithanon-LLMextractivesummarizationbaseline(TextRank).
	\| Method \| \| \| Yelp \| \| AmazonMusic \| \| \| AmazonBooks \| \|
	\| ------ \| --- \| --- \| ---- \| --- \| ----------- \| --- \| --- \| ----------- \| --- \|
	NDCG@1 NDCG@5 NDCG@10 NDCG@1 NDCG@5 NDCG@10 NDCG@1 NDCG@5 NDCG@10
	10H 0.1823 0.2815 0.4928 0.1875 0.3796 0.5153 0.1841 0.3146 0.4263
	KAR(Xietal.,2024) 0.2156 0.3298 0.5412 0.3018 0.4896 0.6015 0.2965 0.4715 0.5834
	RLMRec(Renetal.,2024) 0.2419 0.3472 0.5587 0.3371 0.5434 0.6162 0.2748 0.4526 0.5719
	PALR(Yangetal.,2023) 0.2494 0.3563 0.5691 0.3395 0.5247 0.6115 0.2627 0.4634 0.5538
	LG(Wangetal.,2025) 0.3187 0.4685 0.5814 0.4012 0.5674 0.6489 0.3795 0.5189 0.6284
	R4Rec(Fangetal.,2025) 0.2575 0.3792 0.5526 0.2928 0.5912 0.6343 0.3013 0.4928 0.5959
	Ours 0.3390 0.4873 0.6008 0.5123 0.6165 0.7025 0.4288 0.5638 0.6599
	\| \| Table9: \| RankingperformanceunderEASE-based(Steck,2019)hardnegatives. \| \| \| \| \| \| \| \|
	\| ------------- \| ------- \| ----------------------------------------------------------- \| --- \| --- \| ----------- \| --- \| --- \| ----------- \| --- \|
	\| VarianceGroup \| \| Yelp \| \| \| AmazonMusic \| \| \| AmazonBooks \| \|
	#Samp. MAE RMSE Acc F1 #Samp. MAE RMSE Acc F1 #Samp. MAE RMSE Acc F1
	\| \| \| \| (%) \| (%) \| \| (%) (%) \| \| \| (%) (%) \|
	\| --- \| --- \| --- \| --- \| --- \| --- \| ------- \| --- \| --- \| ------- \|
	Stable 1530 0.4017 0.8664 71.76 68.96 432 0.2368 0.5990 82.56 86.89 1876 0.3171 0.8403 76.64 72.56
	Moderate 1354 0.4773 0.8309 60.09 46.26 436 0.3921 0.6628 64.16 47.02 1827 0.4495 0.8059 66.73 55.45
	Diverse 1444 0.6632 1.1198 51.13 45.95 428 0.5536 0.9607 57.10 59.27 1838 0.6199 1.0603 49.52 46.35
	Table10: PerformanceofDUETunderdifferentlevelsofuserpreferencediversity(measuredbyratingvariance).
	#Samp. denotesthenumberofsamplesineachgroup.