Paper deep dive
Reasoning over Semantic IDs Enhances Generative Recommendation
Yingzhi He, Yan Sun, Junfei Tan, Yuxin Chen, Xiaoyu Kong, Chunxu Shen, Xiang Wang, An Zhang, Tat-Seng Chua
Intelligence
Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 96%
Last extracted: 3/26/2026, 1:45:01 AM
Summary
SIDReasoner is a two-stage framework for generative recommendation that enhances reasoning over Semantic IDs (SIDs) by strengthening SID-language alignment through multi-task training and an enriched synthetic corpus, followed by outcome-driven reinforced optimization (GRPO) to guide reasoning trajectories without explicit annotations.
Entities (5)
Relation Signals (3)
SIDReasoner → improves → Generative Recommendation
confidence 100% · Extensive experiments on three real-world datasets demonstrate the effectiveness of our reasoning-augmented SID-based generative recommendation.
SIDReasoner → uses → GRPO
confidence 100% · we adopt Group Relative Policy Optimization (GRPO) to provide outcome-based feedback
Semantic IDs → isquantizedby → RQ-VAE
confidence 95% · we adopt an item quantization based RQ-VAE
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Recent advances in generative recommendation have leveraged pretrained LLMs by formulating sequential recommendation as autoregressive generation over a unified token space comprising language tokens and itemic identifiers, where each item is represented by a compact sequence of discrete tokens, namely Semantic IDs (SIDs). This SID-based formulation enables efficient decoding over large-scale item corpora and provides a natural interface for LLM-based recommenders to leverage rich world knowledge. Meanwhile, breakthroughs in LLM reasoning motivate reasoning-enhanced recommendation, yet effective reasoning over SIDs remains underexplored and challenging. Itemic tokens are not natively meaningful to LLMs; moreover, recommendation-oriented SID reasoning is hard to evaluate, making high-quality supervision scarce. To address these challenges, we propose SIDReasoner, a two-stage framework that elicits reasoning over SIDs by strengthening SID--language alignment to unlock transferable LLM reasoning, rather than relying on large amounts of recommendation-specific reasoning traces. Concretely, SIDReasoner first enhances SID-language alignment via multi-task training on an enriched SID-centered corpus synthesized by a stronger teacher model, grounding itemic tokens in diverse semantic and behavioral contexts. Building on this enhanced alignment, SIDReasoner further improves recommendation reasoning through outcome-driven reinforced optimization, which guides the model toward effective reasoning trajectories without requiring explicit reasoning annotations. Extensive experiments on three real-world datasets demonstrate the effectiveness of our reasoning-augmented SID-based generative recommendation. Beyond accuracy, the results highlight the broader potential of large reasoning models for generative recommendation, including improved interpretability and cross-domain generalization.
Tags
Links
- Source: https://arxiv.org/abs/2603.23183v1
- Canonical: https://arxiv.org/abs/2603.23183v1
Trouble viewing inline? Open PDF directly →
Full Text
83,473 characters extracted from source content.
Expand or collapse full text
Reasoning over Semantic IDs Enhances Generative Recommendation Yingzhi He heyingzhi@u.nus.edu National University of Singapore Singapore, Singapore Yan Sun sthen@mail.ustc.edu.cn University of Science and Technology of China Hefei, China Junfei Tan sober_clever@mail.ustc.edu.cn University of Science and Technology of China Hefei, China Yuxin Chen yuxin.chen@u.nus.edu National University of Singapore Singapore, Singapore Xiaoyu Kong kongxy@mail.ustc.edu.com University of Science and Technology of China Hefei, China Chunxu Shen ∗ lineshen@tencent.com Tencent Inc. China Xiang Wang xiangwang1223@gmail.com University of Science and Technology of China Hefei, China An Zhang ∗ an.zhang3.14@gmail.com University of Science and Technology of China Hefei, China Tat-Seng Chua dcscts@nus.edu.sg National University of Singapore Singapore, Singapore Abstract Recent advances in generative recommendation have leveraged pretrained LLMs by formulating sequential recommendation as au- toregressive generation over a unified token space comprising lan- guage tokens and itemic identifiers, where each item is represented by a compact sequence of discrete tokens, namely Semantic IDs (SIDs). This SID-based formulation enables efficient decoding over large-scale item corpora and provides a natural interface for LLM- based recommenders to leverage rich world knowledge. Meanwhile, breakthroughs in LLM reasoning motivate reasoning-enhanced rec- ommendation, yet effective reasoning over SIDs remains underex- plored and challenging. Itemic tokens are not natively meaningful to LLMs; moreover, recommendation-oriented SID reasoning is hard to evaluate, making high-quality supervision scarce. To address these challenges, we propose SIDReasoner, a two- stage framework that elicits reasoning over SIDs by strengthen- ing SID–language alignment to unlock transferable LLM reason- ing, rather than relying on large amounts of recommendation- specific reasoning traces. Concretely, SIDReasoner first enhances SID-language alignment via multi-task training on an enriched SID- centered corpus synthesized by a stronger teacher model, grounding itemic tokens in diverse semantic and behavioral contexts. Building ∗ Corresponding Authors. Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for components of this work owned by others than the author(s) must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. Request permissions from permissions@acm.org. Conference’17, Washington, DC, USA © 2026 Copyright held by the owner/author(s). Publication rights licensed to ACM. ACM ISBN 978-x-x-x-x/Y/M https://doi.org/10.1145/n.n on this enhanced alignment, SIDReasoner further improves recom- mendation reasoning through outcome-driven reinforced optimiza- tion, which guides the model toward effective reasoning trajectories without requiring explicit reasoning annotations. Extensive experi- ments on three real-world datasets demonstrate the effectiveness of our reasoning-augmented SID-based generative recommendation. Beyond accuracy, the results highlight the broader potential of large reasoning models for generative recommendation, including im- proved interpretability and cross-domain generalization. Our codes are available at https://github.com/HappyPointer/SIDReasoner. Keywords Recommendation, Large Language Models, Reasoning ACM Reference Format: Yingzhi He, Yan Sun, Junfei Tan, Yuxin Chen, Xiaoyu Kong, Chunxu Shen, Xiang Wang, An Zhang, and Tat-Seng Chua. 2026. Reasoning over Semantic IDs Enhances Generative Recommendation. In . ACM, New York, NY, USA, 16 pages. https://doi.org/10.1145/n.n 1 Introduction Generative recommendation has recently emerged as a promising paradigm for sequential recommendation with an increasing atten- tion on Semantic ID (SID)-based methods [32,40,50,75,76]. By quantizing an item’s semantic embedding into a compact sequence of discrete identifiers called Semantic IDs (SID) [22,40], item recom- mendation can be formulated as autoregressive generation task over SID sequences. This sequence prediction paradigm naturally aligns with the Transformer architecture, facilitating seamless integra- tion of generative methods with pretrained large language models (LLMs), whose rich semantic knowledge can facilitate recommen- dation quality and cross-domain generalization [17,72]. Despite these advances, reasoning, one of the central capabilities behind recent breakthroughs of LLMs [14,25,53,56], remains at an early arXiv:2603.23183v1 [cs.IR] 24 Mar 2026 Conference’17, July 2017, Washington, DC, USAHe et al. stage of exploration in generative recommendation, with its full potential yet to be uncovered. Existing approaches to reasoning in recommendation predomi- nantly represent items with textual descriptions, leveraging the rich world knowledge of LLMs to comprehend item semantics and infer user intentions behind interactions [12,46,48,62,63,71]. These methods can be broadly categorized into two paradigms: explicit reasoning and latent reasoning. Explicit reasoning generates nat- ural language rationales to explicitly articulate user preferences [12,31,48,62,63,66]. The latent reasoning improves efficiency by skipping explicit chain-of-thought generation, but sacrifices interpretability and controllability [11,37,46,71]. However, repre- senting items purely with textual descriptions suffers from severe decoding inefficiency and item grounding issues, limiting its prac- ticality in real-world systems [34]. A recent promising direction avoids such issues by representing items with SIDs and expanding SID tokens into the LLM vocabulary. After the SID-language align- ment training, the model can reason in language and recommend in SIDs within a unified representation space [17,38]. While this paradigm offers new potential for reasoning in recommendation, its effectiveness critically depends on the quality of SID–language alignment, and existing successes largely rely on industrial-scale pretraining with substantial computational cost. In this work, we investigate how to achieve effective and data- efficient reasoning for generative recommendation over SIDs. Rea- soning in recommendation is inherently challenging, as it models implicit user preferences that are not directly observable. This implicitness gives rise to two fundamental challenges: first, high- quality reasoning supervision is scarce and expensive to obtain. Second, the quality of recommendation reasoning is hard to evalu- ate given the implicit nature of user preferences. To tackle these challenges, we propose SIDReasoner, a data- efficient two-stage framework that elicits reasoning over SIDs by strengthening SID–language alignment to unlock transferable LLM reasoning. First, to mitigate the scarcity of high-quality recom- mendation reasoning supervision, we strengthen SID–language alignment as a complement alongside recommendation-specific reasoning traces, enabling general knowledge and reasoning abili- ties to better generalize to recommendation tasks. Concretely, we establish SID–language grounding through multi-task alignment, where the model comprehend semantics of SID tokens through a wide range of recommendation-related data. We further leverage a large teacher model with rich world knowledge to perform se- mantic expansion based on item titles, synthesizing more diverse and informative SID-centered descriptive texts. This enriched cor- pus exposes the model to broader and more fine-grained semantic associations for SID tokens, strengthening their alignment with natural language. Second, once the model attains a comprehensive understanding of SIDs and can transfer general knowledge to rec- ommendation tasks, we enable it to self-explore reasoning patterns that are effective for recommendation. In the absence of a direct evaluation criterion for reasoning quality, we adopt Group Relative Policy Optimization (GRPO) [42] to provide outcome-based feed- back and steer the model toward effective reasoning trajectories. Extensive experiments on three real-world datasets demonstrate that, with enriched SID–language alignment, effective SID align- ment and reasoning can be achieved even at academic-scale dataset, enabling LLMs to perform effective reasoning-based generative rec- ommendation in a data-efficient manner. Beyond effectiveness, our empirical results further reveal the broader potential of reasoning in recommendation, including strong cross-domain generalization and improved interpretability through explicit reasoning processes. We summarize our main contributions as follows: •We investigate how to enable effective reasoning in SID-based generative recommendation, and propose a data-efficient frame- work that allows LLMs to reason over itemic tokens, alleviating the need for explicit reasoning supervision. •We empirically demonstrate the broader potential of reasoning in recommendation beyond accuracy improvements (cf. Section 4.2.1), including cross-domain generalization (cf. Section 3.1) and interpretability (cf. Section 4.4.2). • We design and release an open-sourced reasoning-based genera- tive recommendation pipeline, facilitating future investigations into reasoning-enhanced generative recommendation. 2 Related Work In this section, we briefly review the works related to this paper from two main categories: 1) generative recommendation, and 2) large reasoning models. 2.1 Generative Recommendation With the growing success of LLMs and the transformer architecture, generative recommendation has emerged as a promising research direction in the recommender systems. Existing studies in genera- tive recommendation can be broadly classified into three categories: sparse ID-based methods, text-based methods, and semantic ID- based methods. Sparse ID-based methods represent each item as a unique iden- tifier and serialize user interactions into token sequences, formu- lating recommendation as next-token prediction tasks [1,13,67]. Existing explorationss has evolved along two major directions: incorporating additional features or contexts for controllable gen- eration [1,16,59], and optimizing token structures to improve efficiency [15, 27]. However, sparse ID-based methods suffer from scalability issues due to the large output space and are ineffective in cold-start and long-tail scenarios [34]. Text-based methods represent each item using its natural lan- guage description and generate the textual description of the target item as the recommendation output [8,10,68]. The pretrained lan- guage models, which preserve strong semantic understanding capa- bilities, are further fine-tuned on recommendation tasks [4,6,7,35], exhibiting strong generalization to cold-start items and enabling more interpretable recommendations [33,61]. However, generating long textual sequences also introduces substantial computational overhead and faces inherent challenges in reliably grounding the output to real-world items [5,70], making such approaches difficult to deploy in practical recommendation systems. Semantic ID-based methods represent items with short sequences of discrete codes obtained by quantizing continuous item embed- dings [22,32,40]. Semantic IDs (SIDs) offer a trade-off between semantic expressiveness and decoding efficiency, preserving rich semantic information under a compact decoding space. Subsequent studies further extend this paradigm by integrating collaborative Reasoning over Semantic IDs Enhances Generative RecommendationConference’17, July 2017, Washington, DC, USA LLM SID-Title Translation SID Sequence Prediction The user has interacted with items: Predict the SID of next item. a 3 b 7 c 5 a 1 b 9 c 3 ,, ... Multi-task Fine-tuning a 3 b 7 c 5 by SquareSoft, notable for allowing social engagement. It appeals to player who appreciate dynamic gaming experiences, ... is a landmark RPG developed Item Semantic Enrichment a 3 b 7 c 5 User Reasoning Enrichment The user’s interaction with reveals interest in interactive experiences... ... Thus, he would enjoy cooperative games that emphasize real-time collaboration. Enriched Pre-training Corpus Reasoning Path 1 Reasoning Path 2 Reasoning Path N Score: 0.00 Score: 0.25 Score: 1.00 Reasoning Enhancement with GRPO The user interacted with <a 3 ><b 7 ><c 5 >, <a 1 ><b 9 ><c 3 >, .... Predict the SID of next item. .... <think> The interacted <a 3 ><b 7 ><c 5 > reflects interest ... < > <a 3 ><b 6 ><c 9 > .... .... word tokens SID tokens a 3 b 7 c 5 a 1 b 9 c 3 a 3 b 7 c 5 a 3 b 6 c 9 a x Figure 1: Illustration of the overall framework of our proposed SIDReasoner. signals into SID construction [50,52,57,75,76] and advancing quan- tization schemes to improve representation quality or generation efficiency [24,54,73]. Meanwhile, recent efforts explore the integra- tion of Semantic IDs into pretrained language models by aligning SID tokens with the semantic space of LLMs [17,21,38,72,74]. This enables models to speak in natural language as well as recom- mend by generating SIDs. Such integration unlocks new capabilities of generative recommendation, including interactive recommen- dation, explainable decision-making, and explicit reasoning over user–item interactions [38]. 2.2 Large Reasonging Models In large language models, reasoning decomposes complex prob- lems into a series of simpler steps and allows the model to an- swer with higher confidence through an additional reasoning pro- cess [30,45,49,65]. This is most commonly reflected in Chain-of- Thought (CoT) [53], which introduces explicit token-level inter- mediate steps and trains or prompts the model to solve complex problems by generating these steps prior to the final answer. The quality of reasoning can be further improved by allocating addi- tional inference-time computation, for example through parallel sampling [36,51,58], and iterative refinement [26,28,55]. Moreover, beyond heuristic control at inference time, reinforcement learning with verifiable rewards provides a more direct way to refine the reasoning process, improving both the reliability of intermediate inference and final correctness [2,3,14,25]. Overall, large reason- ing models can be understood as systems that explicitly control and allocate additional test-time computation to achieve stronger problem-solving performance [44, 56, 69]. In recommendation systems, reasoning is similarly introduced to replace one-shot scoring with multi-step inference processes that progressively refine user intent and recommendation decisions. Existing explorations of incorporating reasoning into recommenda- tion tasks can be broadly categorized into two directions: explicit reasoning and latent reasoning. The first direction, explicit reason- ing, relies on explicit natural language rationales as intermediate inference steps, enabling interpretable and controllable multi-step recommendation [12,31,38,48,62,63,66]. In contrast, the sec- ond direction, latent reasoning, performs multi-step inference by iteratively refining the model’s latent states in the continuous rep- resentation space, without relying on explicit natural language rationales [11,37,46,71]. Despite recent progress, both directions still face substantial challenges. Explicit reasoning relies on high- quality recommendation-oriented reasoning data that is not natu- rally available, whereas latent reasoning lacks interpretability and controllability in the decision process. 3 Methodoly In this section, we present our reasoning-enhanced framework for generative recommendation, as illustrated in Figure 1. We first formalize the problem, then describe how we quantize items and align itemic tokens with natural language through multi-task train- ing and enriched semantic corpus. Finally, we present reinforced optimization that enhances the model’s reasoning ability on rec- ommendation. 3.1 Task Formulation 3.1.1Tokenization with Semantic IDs. Generative recommendation aims to generate the next item that a user is likely to interact with. LetUandIdenote the sets of users and items. For each user푢 ∈ U, a chronological interaction history is observed as H 푢 = (푖 1 ,푖 2 , . . .,푖 푇 ), where푖 푡 ∈ I, and the goal is to generate the next item푖 푇+1 conditioned onH 푢 . To leverage a pretrained autoregressive LLM for recommendation, each item푖is represented as a sequence of discrete tokens that can be processed by next- token prediction. Besides the LLM language-token vocabularyV LM , a recommendation-specific itemic token vocabularySis defined, and each item푖 ∈ Iis mapped to a fixed-length Semantic ID (SID) sequence:SID(푖)= (푠 1 푖 ,푠 2 푖 , . . .,푠 퐿 푖 ),푠 푙 푖 ∈ S, where퐿is the Conference’17, July 2017, Washington, DC, USAHe et al. SID length. The SID sequence is obtained by encoding the item metadata푡 푖 (e.g., title, category, and optionally a brief description) into a continuous semantic embedding and then quantizing it into 퐿discrete semantic tokens, yielding a compact tokenization that preserves item semantics. 3.1.2Generative Recommendation with Reasoning. Under this for- mulation, generating the next item푖 푇+1 corresponds to autore- gressively generating its퐿SID tokens fromS, i.e., one item gen- eration is realized by퐿consecutive next-token prediction steps over the itemic vocabulary. The token space of the LLM is aug- mented to include both natural language tokens and itemic tokens, i.e.,V= V LM ∪S, enabling textual reasoning and item genera- tion within a unified autoregressive model. For each user푢, the interaction historyH 푢 = (푖 1 ,푖 2 , . . .,푖 푇 )is converted into itemic token sequences by mapping each interacted item푖 푡 to its SID sequence y 푡 = SID(푖 푡 ) ∈ S 퐿 . The history is then represented as a flattened itemic-token contextH 푢 = concat(y 1 ,y 2 , . . .,y 푇 ). Together with an instruction-style prompt p, the model input is formed asC 푢 = [p;H 푢 ]. Conditioned on this contextC 푢 , the rec- ommender first generates an intermediate reasoning sequence 휏= (푟 1 ,푟 2 , . . .,푟 푀 ), and subsequently generates the next item y 푇+1 = SID(푇 +1). The overall generation process is modeled autoregressively as 휏 ∼ 휋 휃 (· | C 푢 ),y 푇+1 ∼ 휋 휃 (· | C 푢 ,휏),(1) where휋 휃 denotes the LLM-based generative policy. In this formu- lation, the reasoning sequence휏serves as an explicit intermediate inference step that guides the generation of the next item based on the user’s historical behavior. 3.2 Enriched SID-Language Alignment 3.2.1Item Quantization. To obtain discrete Semantic IDs from item textual metadata, we adopt an item quantization based RQ-VAE [64], which has become a widely adopted tokenization method in generative recommendation [40]. Specifically, for each item푖 ∈I, we first encode its textual metadata푡 푖 into a continuous semantic representation z 푖 ∈ R 푑 using a text encoder. The embedding is then fed into a quantization module, where a sequence of discrete codes is produced through a multi-stage residual quantization process. Concretely, we maintain퐿codebooksC 1 ,C 2 , . . .,C 퐿 , where each codebookC 푙 =e 푙 1 , . . .,e 푙 퐾 contains퐾code vectors. At quantiza- tion stage푙, the residual vector r 푙−1 is approximated by selecting the nearest codeword fromC 푙 , and the residual is updated accordingly: 푠 푙 푖 = arg min 푘 ∥r 푙−1 − e 푙 푘 ∥ 2 2 ,r 푙 = r 푙−1 − e 푙 푠 푙 푖 ,(2) with r 0 = z 푖 . After퐿stages, the item embedding is represented as the sum of the selected codewords across all levels. The resulting discrete indices(푠 1 푖 ,푠 2 푖 , . . .,푠 퐿 푖 )constitute the Semantic ID sequence SID(푖)of item푖. The codebooks and the quantization module are optimized with a joint lossL RQ-VAE consisting of a reconstruction termL recon and a residual quantization regularization termL RQ . Specifically, L RQ-VAE =L recon +L RQ ,(3) L recon =∥z 푖 − ˆ z 푖 ∥ 2 2 ,(4) L RQ = 퐿 ∑︁ 푙=1 ∥sg[r 푙−1 ]− e 푙 푠 푙 푖 ∥ 2 2 + 훽∥r 푙−1 − sg[e 푙 푠 푙 푖 ]∥ 2 2 .(5) where ˆ z 푖 denotes the reconstructed embedding obtained from the decoder,sg[·]denotes the stop-gradient operator and훽controls the strength of the commitment term. The residual quantization scheme allows each item to be represented by a compact sequence of discrete SID tokens while preserving semantic fidelity, and enables efficient modeling within an autoregressive language model. 3.2.2 Multi-task Fine-tuning. While SIDs provide a compact dis- crete representation of items, the corresponding SID tokens are newly introduced into the LLM vocabulary with randomly initial- ized embeddings, and therefore are semantically meaningless to the model before fine-tuning. To enable the model to understand, reason over, and generate SID tokens meaningfully in the recommenda- tion context, we conduct a multi-task fine-tuning stage where SID tokens and natural language tokens co-occur across diverse recom- mendation scenarios. Through these tasks, the model jointly learns recommendation signals from SID sequences and the semantic cor- respondence between SID tokens and textual item descriptions, thereby achieving effective SID–language alignment. To this end, we adopt multiple recommendation-oriented tasks under a unified autoregressive objective, which can be grouped into two categories. •Item prediction. Given a user’s historical interactions, the model is trained to predict future interactions, where items are repre- sented with either SID sequences or textual descriptions. This task enables the model to capture behavioral patterns while grounding SID tokens in semantic item information. •SID translation. To facilitate SID-language alignment, the model is further fine-tuned to translate SIDs, where the model generates the textual title from a given SID sequence and, conversely, the SID sequence from its textual title. This task encourages the model to associate SID tokens with their underlying semantic meanings in the language space. Due to space limitations, detailed data examples of each task are provided in appendix C. All tasks are jointly optimized under the standard next-token prediction objective, enabling effective SID–language alignment for generative recommendation. 3.2.3 Enriched Corpus Pre-training. Although the recommender benefits from multi-task training, the limited diversity of task for- mats leads to restricted and repetitive alignment between SID to- kens and word tokens, which results in insufficient understanding of SID semantics and reduces the quality of item-related reason- ing. To address this issue, we leverage a large teacher language model with rich world knowledge to synthesize a more semanti- cally enriched item–language corpus for additional pre-training, as illustrated in Figure 2. The goal is to provide more diverse and informative SID–language associations, thereby strengthening the model’s understanding of SID semantics. In particular, we augment the existing recommendation data from two perspectives. •Item-centric semantic enrichment. For each item, the teacher model first expands its textual metadata into a structured seman- tic description, including attributes like usage scenarios, target users, key characteristics, etc.and then generates a coherent para- graph that interleaves SID tokens with natural language. This Reasoning over Semantic IDs Enhances Generative RecommendationConference’17, July 2017, Washington, DC, USA Structured Instruction Tuning <SID-1>, <SID-2>, ...... à<SID-N> Item-centric semanticenrichment <SID-1> is a story-driven role-playing game on the Nintendo Wii, ...... •<SID-2> allows the Nintendo Wii to connect to the internet ...... ...... •<SID-N> allows more accurate motion trackingfor Nintendo Games. ...... User-centric reasoning enrichment The user’s interactionwith <SID-1> and <SID-2>reflectsstrong interest in Nintendo games ...... Thus, I would recommend <SID-N>. Figure 2: Illustration of enriched alignment corpus. produces training samples where SID tokens appear in semanti- cally rich and linguistically diverse contexts, strengthening their alignment with underlying meanings. • User-centric reasoning enrichment. Similarly, for each user interaction sequence, the teacher model infers the underlying motivations and generates a concise mixed SID–language de- scription that summarizes user’s historical behavior, reasoning over motivations, and the resulting actions. These enriched sam- ples strengthen SID–language alignment and, more importantly, associate SID tokens with user behavior reasoning. In addition, to prevent the model from overfitting to recommen- dation tasks and degrading its general language and reasoning abilities, we further mix the general-domain reasoning data into the alignment corpus. This auxiliary general data helps preserve the model’s original general reasoning ability while learning enriched SID–language associations. 3.3 Reinforced Reasoning Enhancement 3.3.1Cold-start Reasoning Activation. Although the model acquires the ability to jointly reason and recommend through enriched SID–language alignment, it does not necessarily default to generat- ing explicit reasoning before itemic predictions at inference time. To reliably activate the reasoning behavior, we introduce a light- weight cold-start reasoning activation stage. Specifically, we use teacher-generated reasoning from the alignment stage to construct supervised samples, and apply standard supervised fine-tuning to enforce a reason-then-recommend generation pattern, where the model first produces a reasoning sequence휏and then predicts the target SID sequence y 푇+1 conditioned on the user contextC 푢 . It is worth noting that, due to prior SID–language alignment, the model already possesses the capability to perform reasoning-based recommendation. This stage mainly serves to improve the reliability of response formatting, ensuring that reasoning is consistently generated before recommendation. In practice, this activation step is lightweight and takes only a single epoch of fine-tuning. 3.3.2Group-wise Reinforcement Learning. After reasoning activa- tion, the model has a more stable initialization for reasoning in recommendation. We further refine the model’s policy by reinforce- ment learning on recommendation tasks, where the policy휋 휃 is directly optimized with rewards from recommendation accuracy. To this end, we define the reward for a reasoning sequence휏and predicted item representation y as 푅 휃 (휏, y)= 푅 푠푟 (y,푖 푇+1 )+ 휆푅 푓 (y),(6) where휆is a balancing coefficient,푅 푠푟 (·)and푅 푓 (·)are the stepwise rule-based reward and format reward, respectively. The stepwise rule-based reward푅 푠푟 (y,푖 푇+1 )assesses the prediction quality by comparing the generated item representation y with the ground- truth next item. Let퐿be the length of the item representation and 푚be the length of the longest correct prefix between y and the ground-truth y 푇+1 , then 푅 푠푟 (y,푖 푇+1 )= 1 2 퐿−푚 ,(7) which provides a smoothly increasing reward as more tokens in y match the ground truth, approaching 1 when the entire sequence is correct. The format reward푅 푓 (y)checks if the prediction y is structurally valid and corresponds to an existing item: 푅 푓 (y)= ( 1,if y maps to a catalog-existing item, 0,otherwise, (8) which encourages the model to generate valid and meaningful itemic token sequences. With the reward function defined, we optimize the policy휋 휃 us- ing Group Relative Policy Optimization (GRPO) [42]. For each user contextC 푢 , we sample a group of퐾reasoning–prediction trajecto- ries표 푘 퐾 푘=1 from the previous policy휋 휃 old , where each trajectory 표 푘 = 휏 푘 ◦y 푘 consists of a reasoning sequence followed by a SID prediction. Each trajectory is assigned an outcome-based reward푅 푘 . GRPO performs group-wise normalization of rewards and updates the policy by maximizing the following clipped surrogate objective: L GRPO (휃)= E C 푢 " 1 퐾 퐾 ∑︁ 푘=1 min 휌 푘 (휃) ˆ 퐴 푘 , clip(휌 푘 (휃), 1− 휂, 1+휂) ˆ 퐴 푘 # − 훽 퐷 KL 휋 휃 ∥ 휋 ref , (9) where휌 푘 (휃)= 휋 휃 (표 푘 |C 푢 ) 휋 휃 old (표 푘 |C 푢 ) is the trajectory-level importance ratio, ˆ 퐴 푘 denotes the group-normalized advantage from푅 푗 퐾 푗=1 ,휂is the clipping threshold, and 훽 controls KL regularization. By optimizing this objective, the model is encouraged to increase the likelihood of reasoning–prediction trajectories표 푘 that yield higher relative reward, while suppressing less effective reasoning paths. This reinforced optimization progressively refines both the quality of generated reasoning and the accuracy of item prediction. 3.3.3 Reasoning Generalizes to Out-of-Domain Items. 4 Experiments In this section, we present the experimental results and correspond- ing analysis to answer the following research questions (RQs). Conference’17, July 2017, Washington, DC, USAHe et al. Table 1: Performance comparison over different baseline methods. R is shorts for Recall, N is short for NDCG. Models GamesOfficeIndustrial R@5N@5R@10 N@10R@5N@5R@10 N@10R@5N@5R@10 N@10 Caser0.03760.0241 0.0659 0.0332 0.0880 0.0663 0.1114 0.0738 0.0664 0.0528 0.0852 0.0588 GRU4Rec0.03290.0219 0.0599 0.0305 0.0682 0.0480 0.0974 0.0574 0.0788 0.0578 0.1030 0.0649 SASRec0.05010.0345 0.0723 0.0416 0.1019 0.0824 0.1167 0.0871 0.0807 0.0647 0.0964 0.0697 TIGER0.04890.0300 0.0763 0.0402 0.1270 0.1037 0.1429 0.1121 0.1003 0.0823 0.1325 0.0924 HSTU0.05390.0396 0.0746 0.0462 0.1204 0.1069 0.1323 0.1107 0.1008 0.08980.1138 0.0940 LETTER0.04450.0294 0.0709 0.0378 0.13150.10740.15200.11390.10800.0850 0.13890.0950 LCRec0.04410.0274 0.0876 0.0412 0.0964 0.0699 0.1487 0.0867 0.0805 0.0520 0.1330 0.0687 ReaRec0.05680.0381 0.0843 0.0470 0.1173 0.0988 0.1385 0.1057 0.0973 0.0796 0.1205 0.0870 R 2 ec0.0655 0.03990.09310.05250.1147 0.0894 0.1486 0.1004 0.0880 0.0774 0.1253 0.0774 SIDReasoner 0.0710 0.0460 0.1031 0.0563 0.1373 0.1119 0.1648 0.1208 0.1109 0.0905 0.1438 0.1010 GamesOfficeIndustrial 0.04 0.06 0.08 0.10 0.12 0.14 0.16 Recall@10 SID Alignment RL on Games RL on Office RL on Full Data Figure 3: Performance comparison on cross-domain recom- mendation setting. • RQ1: How effective is our proposed SIDReasoner? • RQ2: How does each part contribute to model performance? • RQ3: What are the key properties of SIDReasoner? 4.1 Experiment Settings 4.1.1Datasets and Evaluation. Following prior works [4,5,7], we conduct experiments on three real-world datasets collected from Amazon platform [23]: Video Games (Games), Office Products (Of- fice), and Industrial and Scientific (Industrial). The datasets contain user interaction histories with each interacted item is associated with rich textual metadata. For all datasets, we first apply 5-core filtering to ensure that each user and each item has at least five interactions. To control sequence length and maintain a consistent modeling setting, we truncate each user’s historical interaction sequence using the sliding window with maximum length set to 10. We adopt a temporal split to simulate realistic recommendation scenarios. For each user, interactions are sorted chronologically and divided into training, validation, and test sets with a ratio of 8:1:1, where the most recent interactions are reserved for validation and testing. This time-aware partition ensures that the model is always evaluated on future behaviors. Detailed dataset statistics are presented in Appendix A. For evaluation metrics, we adopt Recall@K and NDCG@K with cutoff values퐾 ∈ 5,10. We follow the full-item ranking setting, where ranking metrics are computed over the entire item set rather than sampled negatives, which is more closely aligned with real- world recommendation scenarios. 4.1.2 Baselines. To validate the effectiveness of our method, we compare SIDReasoner against three different categories of baselines: (1) traditional discriminative sequential recommenders, including Caser [47], GRU4Rec [20], and SASRec [29], which model user be- haviors from interaction sequences in a discriminative manner. (2) generative recommendation methods, including TIGER [40], HSTU [67], LETTER [50], and LCRec [72] (3) Reasoning-based recommen- dation methods, including ReaRec [46] which leverage latent rea- soning, and푅 2 푒푐[62], which incorporate explicit textual reasoning into the recommendation process. Detailed descriptions of each baseline method are presented in Appendix B. 4.1.3 Implementation Details. We adopt Qwen3-1.7B [60] as the backbone model and perform full-parameter fine-tuning through- out all stages. SID tokens are appended to the tokenizer vocabulary with randomly initialized embeddings. During SID–language align- ment training, we use the AdamW optimizer [39] with a batch size of 1024 and apply early stopping based on the model’s performance on predicting the groundtruth SID given historical SID sequences. For enriched corpus construction, we call the GPT4o-mini API to synthesize semantically enriched SID–language data. Detailed prompt design, data examples, and the mixture ratios of different training sources are provided in Appendix C and Appendix D. In the reinforcement learning stage, we implement GRPO based on verl [43]. We set the rollout number to 16, the KL regularization coefficient to 1×10 −3 , and the batch size to 256. The weighting coefficient휆before the format reward is set to 0.1, and the learning rate is fixed at 5× 10 −7 . 4.2 Performance Comparison (RQ1) 4.2.1Reasoning Enhances In-Domain Performance. We present the performance of our proposed SIDReasoner and compare it with other baseline methods on three datasets, with the results recorded Reasoning over Semantic IDs Enhances Generative RecommendationConference’17, July 2017, Washington, DC, USA Figure 4: Ablation performance of different alignment strate- gies with Best-of-푁 reasoning selection. in Table 1. As illustrated by the experiment results, generative rec- ommendation methods generally outperform traditional sequential recommendation approaches across all datasets. This trend con- firms the effectiveness of modeling recommendation as sequence generation over SIDs. Moreover, after reinforcement learning, our method achieves the strongest SID prediction performance, sur- passing not only conventional generative recommenders but also reasoning-enhanced recommendation approaches. These results in- dicate that effective reasoning over SIDs can be successfully learned and exploited to improve recommendation quality. An interesting observation is that the effectiveness of reason- ing varies notably across different item categories, leading to un- even performance gains on different datasets. Specifically, on the Games dataset, where items are rich in semantic information and align well with the world knowledge encoded in LLMs, reasoning- based methods yield substantial improvements over non-reasoning counterparts. In contrast, on the Industrial dataset, where LLMs possess relatively limited domain-relevant knowledge, the benefits brought by reasoning are notably more limited. The effectiveness of reasoning varies across datasets, and this phenomenon is not unique to our method. A highly consistent trend is also observed in R 2 ec, which performs explicit reasoning before recommending. This trend suggests that the effectiveness of reasoning is closely tied to the availability and quality of semantic knowledge that can be leveraged by LLMs. Figure 3 illustrates the cross-domain generalization ability of reasoning-based recommendation. Specifically, we construct a uni- fied SID space covering Games, Office, and Industrial datasets, and perform SID-language alignment on the mixed corpus spanning all three domains. After the model acquires a shared semantic under- standing of SIDs across domains, we conduct reasoning-oriented reinforcement learning on single datasets (Games and Office) as well as full datasets. As shown in Figure 3, RL conducted on a single domain consistently improves reasoning effectiveness on both in- domain and out-of-domain datasets. This observation suggests that the learned reasoning capability is not tied to domain-specific item distributions, but instead captures a more general knowledge of how to reason effectively in recommendation tasks. Such reasoning knowledge, once learned, can be transferred across domains. 4.3 Ablation Study (RQ2) 4.3.1 Alignment Strategies Affect Reasoning Effectiveness. In this subsection, we investigate how different SID-language alignment Table 2: Ablation performance of different alignment strate- gies before and after reinforcement learning optimization. Models GamesOffice R@10N@10R@10N@10 S1 = Multi-task Alignment0.04140.02060.03880.0201 - After RL0.07410.03550.15390.0970 S2 = S1 + Enriched Alignment0.06320.03200.06190.0363 - After RL0.09570.04700.16070.1178 S3 = S2 + General Reasoning0.08060.04500.12310.0883 - After RL0.10310.05630.16480.1208 strategies affect the reasoning potential of recommendation mod- els, with the goal of enabling more effective reasoning on recom- mendation and guiding the model toward a better convergence point after reinforcement learning. To this end, we conduct a sys- tematic comparison of four backbone settings with increasingly diverse training corpora. (1) Vanilla Qwen3-1.7B, where we di- rectly perform reasoning activation via supervised fine-tuning on a pretrained language model, forcing it to learn recommendation reasoning over SID tokens without any explicit alignment stage. (2) Multi-task Alignment, where Qwen3-1.7B is first trained with a multi-task SID–language alignment objective to establish basic se- mantic grounding of SID tokens, followed by reasoning activation. (3) Enriched Alignment, which further augments the multi-task alignment stage with our generated enriched recommendation cor- pus, aiming to expose the model to more diverse and semantically rich SID-language interactions before reasoning activation. (4) En- riched + General Reasoning, where we additionally incorporate general reasoning data into the alignment stage, alleviating cata- strophic forgetting of the model’s general language abilities induced by recommendation-specific optimization. For these four backbone models, we evaluate their behavior from three perspectives: the up- per bound of recommendation reasoning capability, the converged performance after reinforcement learning, and general abilities. Figure 4 presents the Best-of-푁reasoning performance of the four backbone models under different alignment strategies. The hor- izontal axis denotes the number of sampled reasoning trajectories 푁, while the vertical axis reports the corresponding recommenda- tion performance measured by Recall@10. To compute best-of-푁 performance, we randomly sample푁reasoning outputs for each input, select the best reasoning trajectory according to the ground- truth item, and then evaluate the recommendation quality induced by this selected reasoning. The best-of-푁performance reflects the upper bound of model’s reasoning capability. We reports the converged performance of different backbones after RL training in Table 2. The experiment results suggest that models with higher best-of-푁performances consistently converge to better final performance. This alignment between best-of-푁per- formance and post-RL convergence results suggests that Best-of-푁 evaluation serves as an effective indicator of a model’s reason- ing capacity and its optimization potential under reinforcement learning. The results in both Figure 4 and Table 2 indicate that explicit SID–language alignment is a necessary prerequisite for effective reasoning-based recommendation. Pretrained language Conference’17, July 2017, Washington, DC, USAHe et al. Table 3: Performance comparison over general abilities. ModelsMMLU IFEVAL GSM8K Vanilla Qwen3-1.7B0.60850.17930.6850 S1 = Multi-task Alignment0.27600.09060.0060 S2 = S1 + Enriched Alignment0.44640.07390.0330 S3 = S2 + General Reasoning0.55800.14970.5430 Figure 5: Reasoning length and performance change during RL Training on Games Dataset. models lack inherent knowledge of recommendation semantics and item identifiers, making direct reasoning activation insufficient for understanding SID and the recommendation task. Multi-task alignment provides essential semantic grounding, which substan- tially expands the model’s reasoning capacity. Further enriching the SID-language alignment corpus helps LLMs understands the SIDs better, and diversifies the associations between language and itemic tokens. Finally, mixing general reasoning data during alignment alleviates catastrophic forgetting of general language abilities and simultaneously enhances reasoning effectiveness on recommenda- tion tasks. 4.3.2Alignment Strategies Affect General Abilities. Table 3 reports the general language performance of different backbone models un- der various alignment strategies and compares them with the vanilla Qwen3-1.7B model. The evaluated benchmarks include IFEval [77], MMLU [18,19], and GSM8K [9], which assess instruction following, general reasoning, and mathematical reasoning abilities, respec- tively. Overall, training on recommendation tasks leads to inevitable degradation of general language abilities, as recommendation tasks are naturally out-of-domain samples for general language models. This drop in general performance is most pronounced on math- ematical benchmarks, where performance degrades substantially when only multi-task alignment is applied. Enriching the alignment corpus mitigates this degradation but does not prevent it. Incorpo- rating general reasoning data during alignment effectively alleviates catastrophic forgetting, preserving general language abilities that are also important to effective reasoning on recommendation tasks. 4.4 Model Study (RQ3) 4.4.1 Evolution of Reasoning during RL Training. Figure 5 illus- trates the evolution of reasoning behavior during reinforcement learning on the Games dataset. Specifically, we track the average length of generated reasonings (blue curve) alongside the corre- sponding recommendation performance measured by Recall@10 (orange curve) at different training steps. During RL training, the average reasoning length decreases notably in the early stages and converges to a shorter level, whereas recommendation perfor- mance steadily improves. We hypothesize that this phenomenon stems from the nature of the reasoning patterns learned during SID- language alignment. Before RL, the model primarily acquires its reasoning behaviors from samples generated by the teacher model (GPT-4o-mini). While such teacher-generated reasoning provides a useful starting point, it is not necessarily optimized for recommen- dation and often contains redundant or uninformative reasoning patterns that do not contribute to prediction performance. Dur- ing RL, the model rapidly identifies and discards these ineffective reasoning components. As a result, the model progressively shifts toward shorter, more targeted reasoning trajectories that retain essential decision-relevant information while eliminating unnec- essary intermediate steps. Ultimately, the model achieves better recommendation performance with more compact reasoning, indi- cating that effective recommendation reasoning does not require longer chains, but rather more efficient ones. 4.4.2 Case Study. Figure 6 presents a representative example of how explicit reasoning over SIDs enhances generative recommenda- tion. Given the user’s interaction history encoded as SID sequences, the model generates a reasoning trace that first summarizes the user’s key interests from historical interactions, including strategic role-playing games and Nintendo amiibo items. Building on this summary, the model continues thinking and concludes that rec- ommending additional Nintendo amiibo items is more appropriate given the user’s repeated engagement with such products. Conditioned on this reasoning, the model prioritizes SIDs related to Nintendo amiibo, which eventually hits the target item success- fully. This example provides a transparent view into the model’s internal decision-making process, where intermediate reasoning explicitly reveals how user intent is inferred and subsequently used for recommendation. Such reasoning directly shapes the SID decod- ing trajectory and the resulting recommendation list, rather than serving as a post-hoc explanation of the output. 5 Conclusion In this work, we investigate how to enable effective reasoning in SID-based generative recommendation. We propose SIDReasoner, a two-stage framework that effectively allows large language models to reason over itemic tokens. Stronger SID–language alignment can help general reasoning abilities better generalize to recommen- dation tasks. Motivated by this insight, our approach leverages multi-task alignment and teacher-assisted semantic expansion to establish strong semantic grounding between SID and language tokens. Building on this foundation, outcome-driven reinforcement further guides the model to self-explore effective reasoning pat- terns in recommendation. Extensive experiments on real-world datasets demonstrate that incorporating reasoning into generative recommendation yields competitive performance, together with improved cross-domain generalization and interpretability. Reasoning over Semantic IDs Enhances Generative RecommendationConference’17, July 2017, Washington, DC, USA <a_162><b_125><c_53> <a_162><b_122><c_56> <a_162><b_106><c_98> ...... Nintendo amiibo series collectible games Strategic role-play games Reasoning: Input: The user has sequentially interacted with items <a_1><b_25><c_254>, <a_1><b_25><c_194>, <a_211><b_159><c_123>, <a_211><b_159><c_71>, <a_113><b_159><c_2>, <a_162><b_174><c_61>, <a_162><b_214><c_137> Can you recommend the next item for him? Let's think step by step before making recommendation... Analyzing the user's interaction history, I notice a strong inclination towards RPGs, particularly those withinthe Fire Emblem series,as evidenced by their engagement with <a_1><b_25><c_254>and <a_1><b_25><c_194>. This suggests a preference for strategic gameplay and character-driven narratives. Additionally, the user has showninterest in amiibo figures,specifically those from the Super Smash Bros series, as indicated by<a_162><b_174><c_61>and<a_162><b_214><c_137>. This indicates a desire for collectibles that enhance gameplay experiences, likely driven by nostalgia and the desire to unlock additional content in games. The user's consistent engagement withNintendo titles and amiibo suggests they may enjoyitems that not only serve as collectibles but also enhance their gaming experience. Given this pattern, it's reasonable to predict that they may be interested in further amiibo figuresor additional items that complement their existing collectionparticularly those that enhance gameplay or provide unique in-game benefits. Title: Reflet amiibo- Japan Import Output A captivating collectible figure from the Super Smash Bros. series ... for Nintendo enthusiasts... Description: Figure 6: Case study of how explicit reasoning yields effective and explainable recommendations. For future work, we plan to explore whether reasoning over SIDs yields further gains with larger backbone models, and to investi- gate the effectiveness of enriched SID–language alignment under substantially larger-scale training data. Due to limited time and computational resources, we leave scaling to larger models and datasets for future exploration. References [1] Prabhat Agarwal, Anirudhan Badrinath, Laksh Bhasin, Jaewon Yang, Edoardo Botta, Jiajing Xu, and Charles Rosenberg. 2025. Pinrec: Outcome-conditioned, multi-token generative retrieval for industry-scale recommendation systems. arXiv preprint arXiv:2504.10507 (2025). [2] Pranjal Aggarwal and Sean Welleck. 2025. L1: Controlling how long a reasoning model thinks with reinforcement learning. arXiv preprint arXiv:2503.04697 (2025). [3]Daman Arora and Andrea Zanette. 2025. Training language models to reason efficiently. arXiv preprint arXiv:2502.04463 (2025). [4] Keqin Bao, Jizhi Zhang, Wenjie Wang, Yang Zhang, Zhengyi Yang, Yanchen Luo, Chong Chen, Fuli Feng, and Qi Tian. 2025. A bi-step grounding paradigm for large language models in recommendation systems. ACM Transactions on Recommender Systems 3, 4 (2025), 1–27. [5] Keqin Bao, Jizhi Zhang, Yang Zhang, Xinyue Huo, Chong Chen, and Fuli Feng. 2024. Decoding matters: Addressing amplification bias and homogeneity issue in recommendations for large language models. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 10540–10552. [6]Keqin Bao, Jizhi Zhang, Yang Zhang, Wenjie Wang, Fuli Feng, and Xiangnan He. 2023. Tallrec: An effective and efficient tuning framework to align large language model with recommendation. In Proceedings of the 17th ACM conference on recommender systems. 1007–1014. [7] Yuxin Chen, Junfei Tan, An Zhang, Zhengyi Yang, Leheng Sheng, Enzhi Zhang, Xiang Wang, and Tat-Seng Chua. 2024. On softmax direct preference optimization for recommendation. Advances in Neural Information Processing Systems 37 (2024), 27463–27489. [8] Zhixuan Chu, Hongyan Hao, Xin Ouyang, Simeng Wang, Yan Wang, Yue Shen, Jinjie Gu, Qing Cui, Longfei Li, Siqiao Xue, et al.2023. Leveraging large language models for pre-trained recommender systems. arXiv preprint arXiv:2308.10837 (2023). [9]Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Jacob Hilton, Reiichiro Nakano, Christopher Hesse, and John Schulman. 2021. Training Verifiers to Solve Math Word Problems. arXiv:2110.14168 [cs.LG] [10] Zeyu Cui, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. 2022. M6-rec: Generative pretrained language models are open-ended recommender systems. arXiv preprint arXiv:2205.08084 (2022). [11]Sunhao Dai, Jiakai Tang, Jiahua Wu, Kun Wang, Yuxuan Zhu, Bingjun Chen, Bangyang Hong, Yu Zhao, Cong Fu, Kangle Wu, et al.2025. Onepiece: Bringing context engineering and reasoning to industrial cascade ranking system. arXiv preprint arXiv:2509.18091 (2025). [12] Yi Fang, Wenjie Wang, Yang Zhang, Fengbin Zhu, Qifan Wang, Fuli Feng, and Xiangnan He. 2025. Reason4Rec: Large Language Models for Recommendation with Deliberative User Preference Alignment. arXiv preprint arXiv:2502.02061 (2025). [13]Shijie Geng, Shuchang Liu, Zuohui Fu, Yingqiang Ge, and Yongfeng Zhang. 2022. Recommendation as language processing (rlp): A unified pretrain, personalized prompt & predict paradigm (p5). In Proceedings of the 16th ACM conference on recommender systems. 299–315. [14]Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al.2025. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948 (2025). [15]Hao Guo, Erpeng Xue, Lei Huang, Shichao Wang, Xiaolei Wang, Lei Wang, Jin- peng Wang, and Sheng Chen. 2025. Action is All You Need: Dual-Flow Generative Ranking Network for Recommendation. arXiv preprint arXiv:2505.16752 (2025). [16]Ruidong Han, Bin Yin, Shangyu Chen, He Jiang, Fei Jiang, Xiang Li, Chi Ma, Mincong Huang, Xiaoguang Li, Chunzhen Jing, et al.2025. Mtgr: Industrial- scale generative recommendation framework in meituan. In Proceedings of the 34th ACM International Conference on Information and Knowledge Management. 5731–5738. [17]Ruining He, Lukasz Heldt, Lichan Hong, Raghunandan Keshavan, Shifan Mao, Nikhil Mehta, Zhengyang Su, Alicia Tsai, Yueqi Wang, Shao-Chuan Wang, et al. 2025. Plum: Adapting pre-trained language models for industrial-scale generative Conference’17, July 2017, Washington, DC, USAHe et al. recommendations. arXiv preprint arXiv:2510.07784 (2025). [18]Dan Hendrycks, Collin Burns, Steven Basart, Andrew Critch, Jerry Li, Dawn Song, and Jacob Steinhardt. 2021. Aligning AI With Shared Human Values. Proceedings of the International Conference on Learning Representations (ICLR) (2021). [19]Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. 2021. Measuring Massive Multitask Language Under- standing. Proceedings of the International Conference on Learning Representations (ICLR) (2021). [20] Balázs Hidasi and Alexandros Karatzoglou. 2018. Recurrent neural networks with top-k gains for session-based recommendations. In CIKM. 843–852. [21]Minjie Hong, Yan Xia, Zehan Wang, Jieming Zhu, Ye Wang, Sihang Cai, Xi- aoda Yang, Quanyu Dai, Zhenhua Dong, Zhimeng Zhang, et al.2025. EAGER- LLM: Enhancing Large Language Models as Recommenders through Exogenous Behavior-Semantic Integration. In Proceedings of the ACM on Web Conference 2025. 2754–2762. [22]Yupeng Hou, Zhankui He, Julian McAuley, and Wayne Xin Zhao. 2023. Learning vector-quantized item representation for transferable sequential recommenders. In Proceedings of the ACM Web Conference 2023. 1162–1171. [23]Yupeng Hou, Jiacheng Li, Zhankui He, An Yan, Xiusi Chen, and Julian McAuley. 2024. Bridging language and items for retrieval and recommendation. arXiv preprint arXiv:2403.03952 (2024). [24] Yupeng Hou, Jiacheng Li, Ashley Shin, Jinsung Jeon, Abhishek Santhanam, Wei Shao, Kaveh Hassani, Ning Yao, and Julian McAuley. 2025. Generating long semantic ids in parallel for recommendation. In Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2. 956–966. [25] Zhenyu Hou, Xin Lv, Rui Lu, Jiajie Zhang, Yujiang Li, Zijun Yao, Juanzi Li, Jie Tang, and Yuxiao Dong. 2025. Advancing language model reasoning through reinforcement learning and inference scaling. arXiv preprint arXiv:2501.11651 (2025). [26] Jie Huang, Xinyun Chen, Swaroop Mishra, Huaixiu Steven Zheng, Adams Wei Yu, Xinying Song, and Denny Zhou. 2023. Large language models cannot self-correct reasoning yet. arXiv preprint arXiv:2310.01798 (2023). [27]Yanhua Huang, Yuqi Chen, Xiong Cao, Rui Yang, Mingliang Qi, Yinghao Zhu, Qingchang Han, Yaowei Liu, Zhaoyu Liu, Xuefeng Yao, et al.2025. Towards Large-scale Generative Ranking. arXiv preprint arXiv:2505.04180 (2025). [28]Ryo Kamoi, Yusen Zhang, Nan Zhang, Jiawei Han, and Rui Zhang. 2024. When can llms actually correct their own mistakes? a critical survey of self-correction of llms. Transactions of the Association for Computational Linguistics 12 (2024), 1417–1440. [29]Wang-Cheng Kang and Julian McAuley. 2018. Self-attentive sequential recom- mendation. In ICDM. IEEE, 197–206. [30]Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. 2022. Large language models are zero-shot reasoners. Advances in neural information processing systems 35 (2022), 22199–22213. [31]Xiaoyu Kong, Junguang Jiang, Bin Liu, Ziru Xu, Han Zhu, Jian Xu, Bo Zheng, Jiancan Wu, and Xiang Wang. 2025.Think before Recommendation: Au- tonomous Reasoning-enhanced Recommender. CoRR abs/2510.23077 (2025). arXiv:2510.23077 doi:10.48550/ARXIV.2510.23077 [32] Xiaoyu Kong, Leheng Sheng, Junfei Tan, Yuxin Chen, Jiancan Wu, An Zhang, Xiang Wang, and Xiangnan He. 2025. MiniOneRec: An Open-Source Framework for Scaling Generative Recommendation. arXiv preprint arXiv:2510.24431 (2025). [33]Yuxuan Lei, Jianxun Lian, Jing Yao, Xu Huang, Defu Lian, and Xing Xie. 2024. Recexplainer: Aligning large language models for explaining recommendation models. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discov- ery and Data Mining. 1530–1541. [34] Xiaopeng Li, Bo Chen, Junda She, Shiteng Cao, You Wang, Qinlin Jia, Haiying He, Zheli Zhou, Zhao Liu, Ji Liu, et al.2025. A survey of generative recommendation from a tri-decoupled perspective: Tokenization, architecture, and optimization. (2025). [35]Jiayi Liao, Sihang Li, Zhengyi Yang, Jiancan Wu, Yancheng Yuan, Xiang Wang, and Xiangnan He. 2024. Llara: Large language-recommendation assistant. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval. 1785–1795. [36]Hunter Lightman, Vineet Kosaraju, Yura Burda, Harri Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. 2023. Let’s Verify Step by Step. arXiv preprint arXiv:2305.20050 (2023). [37]Xinyu Lin, Haihan Shi, Wenjie Wang, Fuli Feng, Qifan Wang, See-Kiong Ng, and Tat-Seng Chua. 2025. Order-agnostic identifier for large language model-based generative recommendation. In Proceedings of the 48th international ACM SIGIR conference on research and development in information retrieval. 1923–1933. [38]Zhanyu Liu, Shiyao Wang, Xingmei Wang, Rongzhou Zhang, Jiaxin Deng, Honghui Bao, Jinghao Zhang, Wuchao Li, Pengfei Zheng, Xiangyu Wu, et al. 2025. Onerec-think: In-text reasoning for generative recommendation. arXiv preprint arXiv:2510.11639 (2025). [39]Ilya Loshchilov and Frank Hutter. 2017. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101 (2017). [40] Shashank Rajput, Nikhil Mehta, Anima Singh, Raghunandan Hulikal Keshavan, Trung Vu, Lukasz Heldt, Lichan Hong, Yi Tay, Vinh Tran, Jonah Samost, et al. 2023. Recommender systems with generative retrieval. Advances in Neural Information Processing Systems 36 (2023), 10299–10315. [41]John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. 2017. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347 (2017). [42]Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Yang Wu, et al.2024. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300 (2024). [43]Guangming Sheng, Chi Zhang, Zilingfeng Ye, Xibin Wu, Wang Zhang, Ru Zhang, Yanghua Peng, Haibin Lin, and Chuan Wu. 2025. Hybridflow: A flexible and efficient rlhf framework. In Proceedings of the Twentieth European Conference on Computer Systems. 1279–1297. [44]Leheng Sheng, An Zhang, Zijian Wu, Weixiang Zhao, Changshuo Shen, Yi Zhang, Xiang Wang, and Tat-Seng Chua. 2025. On Reasoning Strength Planning in Large Reasoning Models. arXiv preprint arXiv:2506.08390 (2025). [45]Charlie Snell, Jaehoon Lee, Kelvin Xu, and Aviral Kumar. 2024. Scaling llm test- time compute optimally can be more effective than scaling model parameters. arXiv preprint arXiv:2408.03314 (2024). [46]Jiakai Tang, Sunhao Dai, Teng Shi, Jun Xu, Xu Chen, Wen Chen, Jian Wu, and Yuning Jiang. 2025. Think before recommend: Unleashing the latent reasoning power for sequential recommendation. arXiv preprint arXiv:2503.22675 (2025). [47]Jiaxi Tang and Ke Wang. 2018. Personalized top-n sequential recommenda- tion via convolutional sequence embedding. In Proceedings of the eleventh ACM international conference on web search and data mining. 565–573. [48] Alicia Tsai, Adam Kraft, Long Jin, Chenwei Cai, Anahita Hosseini, Taibai Xu, Zemin Zhang, Lichan Hong, Ed H Chi, and Xinyang Yi. 2024. Leveraging llm reasoning enhances personalized recommender systems. In Findings of the Asso- ciation for Computational Linguistics: ACL 2024. 13176–13188. [49] Shenzhi Wang, Le Yu, Chang Gao, Chujie Zheng, Shixuan Liu, Rui Lu, Kai Dang, Xionghui Chen, Jianxin Yang, Zhenru Zhang, et al.2025. Beyond the 80/20 rule: High-entropy minority tokens drive effective reinforcement learning for llm reasoning. arXiv preprint arXiv:2506.01939 (2025). [50]Wenjie Wang, Honghui Bao, Xinyu Lin, Jizhi Zhang, Yongqi Li, Fuli Feng, See- Kiong Ng, and Tat-Seng Chua. 2024. Learnable item tokenization for generative recommendation. In Proceedings of the 33rd ACM International Conference on Information and Knowledge Management. 2400–2409. [51]Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. 2022. Self-consistency improves chain of thought reasoning in language models. arXiv preprint arXiv:2203.11171 (2022). [52]Ye Wang, Jiahao Xun, Minjie Hong, Jieming Zhu, Tao Jin, Wang Lin, Haoyuan Li, Linjun Li, Yan Xia, Zhou Zhao, et al.2024. Eager: Two-stream generative recommender with behavior-semantic collaboration. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 3245–3254. [53]Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al.2022. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems 35 (2022), 24824–24837. [54] Tianxin Wei, Xuying Ning, Xuxing Chen, Ruizhong Qiu, Yupeng Hou, Yan Xie, Shuang Yang, Zhigang Hua, and Jingrui He. 2025. CoFiRec: Coarse-to-Fine Tokenization for Generative Recommendation. arXiv preprint arXiv:2511.22707 (2025). [55]Sean Welleck, Ximing Lu, Peter West, Faeze Brahman, Tianxiao Shen, Daniel Khashabi, and Yejin Choi. 2022. Generating sequences by learning to self-correct. arXiv preprint arXiv:2211.00053 (2022). [56] Yangzhen Wu, Zhiqing Sun, Shanda Li, Sean Welleck, and Yiming Yang. 2025. Inference scaling laws: An empirical analysis of compute-optimal inference for LLM problem-solving. In The Thirteenth International Conference on Learning Representations. [57]Longtao Xiao, Haozhao Wang, Cheng Wang, Linfei Ji, Yifan Wang, Jieming Zhu, Zhenhua Dong, Rui Zhang, and Ruixuan Li. 2025. Unger: Generative recommendation with a unified code via semantic and collaborative integration. ACM Transactions on Information Systems 44, 2 (2025), 1–31. [58]Yuxi Xie, Kenji Kawaguchi, Yiran Zhao, James Xu Zhao, Min-Yen Kan, Junxian He, and Michael Xie. 2023. Self-evaluation guided beam search for reasoning. Advances in Neural Information Processing Systems 36 (2023), 41618–41650. [59]Bencheng Yan, Shilei Liu, Zhiyuan Zeng, Zihao Wang, Yizhen Zhang, Yujin Yuan, Langming Liu, Jiaqi Liu, Di Wang, Wenbo Su, et al.2025. Unlocking Scaling Law in Industrial Recommendation Systems with a Three-step Paradigm based Large User Model. arXiv preprint arXiv:2502.08309 (2025). [60]An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al.2025. Qwen3 technical report. arXiv preprint arXiv:2505.09388 (2025). [61]Zhengyi Yang, Jiancan Wu, Yanchen Luo, Jizhi Zhang, Yancheng Yuan, An Zhang, Xiang Wang, and Xiangnan He. 2023. Large language model can interpret latent space of sequential recommender. ACM Transactions on Information Systems (2023). Reasoning over Semantic IDs Enhances Generative RecommendationConference’17, July 2017, Washington, DC, USA [62]Runyang You, Yongqi Li, Xinyu Lin, Xin Zhang, Wenjie Wang, Wenjie Li, and Liqiang Nie. 2025. R 2 ec: Towards Large Recommender Models with Reasoning. CoRR abs/2505.16994 (2025). arXiv:2505.16994 doi:10.48550/ARXIV.2505.16994 [63]Qihang Yu, Kairui Fu, Shengyu Zhang, Zheqi Lv, Fan Wu, and Fei Wu. 2025. ThinkRec: Thinking-based recommendation via LLM. arXiv preprint arXiv:2505.15091 (2025). [64] Neil Zeghidour, Alejandro Luebs, Ahmed Omran, Jan Skoglund, and Marco Tagliasacchi. 2021. Soundstream: An end-to-end neural audio codec. IEEE/ACM Transactions on Audio, Speech, and Language Processing 30 (2021), 495–507. [65]Eric Zelikman, Yuhuai Wu, Jesse Mu, and Noah Goodman. 2022. Star: Boot- strapping reasoning with reasoning. Advances in Neural Information Processing Systems 35 (2022), 15476–15488. [66]Ziyang Zeng, Heming Jing, Jindong Chen, Xiangli Li, Hongyu Liu, Yixuan He, Zhengyu Li, Yige Sun, Zheyong Xie, Yuqing Yang, et al.2025. Optimizing Gen- erative Ranking Relevance via Reinforcement Learning in Xiaohongshu Search. arXiv preprint arXiv:2512.00968 (2025). [67] Jiaqi Zhai, Lucy Liao, Xing Liu, Yueming Wang, Rui Li, Xuan Cao, Leon Gao, Zhaojie Gong, Fangda Gu, Jiayuan He, et al.2024. Actions speak louder than words: trillion-parameter sequential transducers for generative recommendations. In Proceedings of the 41st International Conference on Machine Learning. 58484– 58509. [68] Junjie Zhang, Ruobing Xie, Yupeng Hou, Xin Zhao, Leyu Lin, and Ji-Rong Wen. 2025. Recommendation as instruction following: A large language model em- powered recommendation approach. ACM Transactions on Information Systems 43, 5 (2025), 1–37. [69] Qiyuan Zhang, Fuyuan Lyu, Zexu Sun, Lei Wang, Weixu Zhang, Zhihan Guo, Yufei Wang, Irwin King, Xue Liu, and Chen Ma. 2025. What, how, where, and how well? a survey on test-time scaling in large language models. CoRR (2025). [70]Wenlin Zhang, Chuhan Wu, Xiangyang Li, Yuhao Wang, Kuicai Dong, Yichao Wang, Xinyi Dai, Xiangyu Zhao, Huifeng Guo, and Ruiming Tang. 2025. Llmtreerec: Unleashing the power of large language models for cold-start recom- mendations. In Proceedings of the 31st International Conference on Computational Linguistics. 886–896. [71]Yang Zhang, Wenxin Xu, Xiaoyan Zhao, Wenjie Wang, Fuli Feng, Xiangnan He, and Tat-Seng Chua. 2025. Reinforced Latent Reasoning for LLM-based Recommendation. arXiv preprint arXiv:2505.19092 (2025). [72] Bowen Zheng, Yupeng Hou, Hongyu Lu, Yu Chen, Wayne Xin Zhao, Ming Chen, and Ji-Rong Wen. 2024. Adapting large language models by integrating collaborative semantics for recommendation. In 2024 IEEE 40th International Conference on Data Engineering (ICDE). IEEE, 1435–1448. [73] Qiyong Zhong, Jiajie Su, Yunshan Ma, Julian McAuley, and Yupeng Hou. 2025. Pctx: Tokenizing Personalized Context for Generative Recommendation. arXiv preprint arXiv:2510.21276 (2025). [74]Guorui Zhou, Honghui Bao, Jiaming Huang, Jiaxin Deng, Jinghao Zhang, Junda She, Kuo Cai, Lejian Ren, Lu Ren, Qiang Luo, et al.2025. OpenOneRec Technical Report. arXiv preprint arXiv:2512.24762 (2025). [75]Guorui Zhou, Jiaxin Deng, Jinghao Zhang, Kuo Cai, Lejian Ren, Qiang Luo, Qian- qian Wang, Qigen Hu, Rui Huang, Shiyao Wang, et al.2025. OneRec Technical Report. arXiv preprint arXiv:2506.13695 (2025). [76] Guorui Zhou, Hengrui Hu, Hongtao Cheng, Huanjie Wang, Jiaxin Deng, Jinghao Zhang, Kuo Cai, Lejian Ren, Lu Ren, Liao Yu, et al.2025. Onerec-v2 technical report. arXiv preprint arXiv:2508.20900 (2025). [77]Jeffrey Zhou, Tianjian Lu, Swaroop Mishra, Siddhartha Brahma, Sujoy Basu, Yi Luan, Denny Zhou, and Le Hou. 2023. Instruction-Following Evaluation for Large Language Models. arXiv preprint arXiv:2311.07911 (2023). Conference’17, July 2017, Washington, DC, USAHe et al. Table 4: The statistics of recommendation datasets. Dataset #Items #Interactions Train Val&Test Games3,85861,41749,1336,142 Office3,45948,65638,9244,866 Industrial3,68645,32536,2594,533 A Dataset Statistics All our recommendation datasets are downloaded from Amazon18 dataset platform and cover three categories: Video Games, Office Products, and Industrial and Scientific. Detailed statistics of each selected category are present in Table 4. B Baseline Methods This section provides a brief introduction to the baselines used in our experiments. • Caser [47] models sequential user behaviors by applying con- volutional neural networks over user interaction sequences to capture local and global sequential patterns. It is trained to pre- dict the next item based on the user’s historical interactions using the binary cross-entropy loss. •GRU4Rec [20] utilizes GRU modules to capture sequential de- pendencies within user interaction sequences. It is trained to predict the next item in a user’s sequence based on previously purchased items. We use the binary cross-entropy loss as opti- mization objective during the training process. •SASRec [29] is a transformer-based recommender widely used in sequential recommendation. It leverages self-attention to cap- ture long-range dependencies in user interaction sequences, en- hancing the accuracy of future interaction predictions. Binary cross-entropy loss is used as the optimization objective. •TIGER [40] is a representative generative recommender that represents items as Semantic IDs and utilizes a Transformer model to autoregressively predict the Semantic IDs (SIDs) of next item. The model is optimized with cross-entropy loss. • HSTU [67] follows the formulation of TIGER, and models the recommendation task as sequence prediction over action and item tokens. Cross-entropy loss is used during model training. •LETTER [50] follows the structure of TIGER and further in- tegrate collaborative signals into SIDs during the quantization stage. The model is optimized using the cross-entropy loss. • LCRec [72] adapts a pre-trained LLMs as the backbone of the gen- erative recommendation. It further enhances SID understanding via a series of recommendation-related tasks. For fair comparison, we implement LCRec by with exactly same training inputs as our multi-task alignment. •ReaRec [46] enhances the SASRec model with latent reasoning, and boosts recommendation effectiveness by introducing extra test-time compute. We use the Progressive Reasoning Learning method cross-Entropy loss. •R 2 ec [62] is constructed based on a pre-trained LLM. Given the historical interactions described in natural language, it leverages explicit textual reasoning before the recommendation decision. The model is trained using Proximal Policy Optimization (PPO) [41] with recommendation-specific rewards. C Alignment Task Formats In this section, we present the templates used for the alignment tasks, which are designed to bridge the gap between textual item information and their corresponding Semantic IDs. Cases 1-2: SID–Title Translation: This category involves bi- directional mapping between textual item titles and their corre- sponding SIDs. Cases 3–6: Generative Next-Item Prediction: These tasks re- quire the model to predict future interactions by alternating input and output representations between item titles and SIDs, which enables the model to capture sequential behavioral patterns while reinforcing the correspondence between SIDs and textual descrip- tions. Case 7: Alignment via Item-centric Semantic Enrichment: This task requires the model to learn to associate SIDs with detailed, augmented item contexts by integrating semantic identifiers into LLM-generated comprehensive descriptions. Case 8: Sequential Recommendation via User-centric Rea- soning Augmentation: This task is to train the model on hybrid narratives of interaction history that interleave Natural Language with Semantic IDs. The goal is to let the model understand user behavior or recommendation logic as a coherent story. Case 1 & 2: SID translation Title to SID <|im_start|>system Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request. Answer the question about item identification.<|im_end|> <|im_start|>user Which item has the title: "Final Fantasy VIII"?<|im_end|> <|im_start|>assistant <a_195><b_133><c_138><|im_end|> SID to Title <|im_start|>system Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request. Answer the question about item identification.<|im_end|> <|im_start|>user What is the title of item <a_175><b_83><c_8>?<|im_end|> <|im_start|>assistant "The Legend of Zelda: Spirit Tracks"<|im_end|> Reasoning over Semantic IDs Enhances Generative RecommendationConference’17, July 2017, Washington, DC, USA Case 3: Item prediction SID history to Title prediction <|im_start|>system Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.Can you recommend the next item for the user based on their interaction history?<|im_end|> <|im_start|>user The user has sequentially interacted with items <a_45><b_88><c_206>, <a_205><b_247><c_168>, <a_45><b_190><c_242>. Can you recommend the next item for him? Tell me the title of the item?<|im_end|> <|im_start|>assistant "PlayStation 4 500GB Console [Old Model][Discontinued]"<|im_end|> Case 4: Item prediction Title prediction <|im_start|>system Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.Given a list of games the user recently enjoy, please write a new game that the user may bought.<|im_end|> <|im_start|>user The user has played the following games before:"FIFA Soccer 12 - Playstation 3", "PS3 Gran Turismo 5 XL Edition", "MLB 13 The Show - Playstation 3"?<|im_end|> <|im_start|>assistant "The Legend of Zelda: Spirit Tracks"<|im_end|> Case 5: Item prediction SID history to SID prediction <|im_start|>system Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request. Can you predict the next possible item that the user may expect?<|im_end|> <|im_start|>user The user has interacted with items <a_45><b_88><c_206>, <a_205><b_247><c_168>, <a_45><b_190><c_242> in chronological order. Can you predict the next possible item that the user may expect? <|im_end|> <|im_start|>assistant <a_231><b_28><c_63><|im_end|> Case 6: Item prediction Title history to SID prediction <|im_start|>system Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request. Based on the user’s historical interaction with item titles, predict the semantic ID of the next item they may expect.<|im_end|> <|im_start|>user The user has interacted with the following games items in chronological order: "FIFA Soccer 12 - Playstation 3", "PS3 Gran Turismo 5 XL Edition", "MLB 13 The Show - Playstation 3". Can you predict the next item the user may expect? <|im_end|> <|im_start|>assistant <a_231><b_28><c_63><|im_end|> D Prompts for Corpus Enrichment We provide the complete instructions for the prompts used in Item- centric Semantic Enrichment and User-centric Reasoning Enrich- ment below, both of which follow a two-stage framework. Item-centric Semantic Enrichment Stage 1: We require the model to reason through raw metadata to generate structured insights, including target audiences, use cases, and latent keywords, transforming sparse data into deep contex- tual knowledge. Stage 2: This stage enforces fusing all metadata into a single paragraph by seamlessly interleaving SIDs into the text. User-centric Reasoning Enrichment Stage 1 focuses on Reasoning Generation, where the model adopts an analyst’s persona to infer latent user preferences and behavioral shifts from interaction history. Stage 2 then performs Narrative Integration, merging the raw interaction sequence with the gener- ated reasoning into a seamless, natural-language paragraph that strictly uses Semantic IDs to represent items. Conference’17, July 2017, Washington, DC, USAHe et al. Case 7: SID translation Interleaving SIDs with Augmented Item Contexts Introducing <a_175><b_24><c_4>, an exhilarating platforming game developed by Nintendo that revolutionizes the beloved Super Mario franchise through its introduction of simultaneous multiplayer action. For the first time ever, players can join forces or compete against one another as they traverse iconic side-scrolling environments brimming with nostalgic charm and fresh challenges. The game supports 2 to 4 players, making it an ideal choice for family game nights where everyone can engage in either cooperative or competitive gameplay. Additionally, casual gaming sessions among friends seeking an interactive and entertaining experience are perfect for immersing in the whimsical world of Mario. The single-player experience also caters to long-time fans craving the classic Mario gameplay they know and love, enhanced with updated graphics and features that breathe new life into the franchise. With characters such as Mario, Luigi, and two Toads, alongside exciting new power-ups like the propeller suit and Penguin Mario, <a_175><b_24><c_4> combines nostalgia with innovative gameplay mechanics, making it a standout title in the Mario series. The intuitive controls of the Wii Remote allow players to navigate with ease, offering both classic and modern input options that appeal to diverse gaming preferences. Players can lift each other to safety or strategically toss opponents into danger, fostering a playful yet competitive environment that is sure to delight. As players race to complete levels, they are ranked at the end based on their scores, coins collected, and enemies defeated, adding an extra layer of friendly rivalry. <a_175><b_24><c_4> truly encapsulates the essence of family entertainment, delivering a multiplayer platformer experience that resonates with children, families, and fans alike while ensuring a fun-filled adventure steeped in the rich legacy of the Super Mario franchise. Case 8: Item prediction Narrative-based Sequential Recommendation with Semantic IDs The user has engaged with <a_250><b_39><c_2>, <a_21><b_19><c_204>, <a_92><b_90><c_145>, and <a_21><b_19><c_204>, showcasing a distinctive pattern in their gaming preferences. This journey began with an intriguing foray into the world of immersive RPGs, starting with <a_250><b_39><c_2>, a title that reflects their appreciation for vibrant storytelling and character development. Following this initial adventure, the user showed a keen interest in enhancing their gaming experience through <a_21><b_19><c_204>, engaging with this technical accessory not once but twice.This indicates not only a desire for improved online gaming capabilities but also a preference for stable performance during their gaming sessions. The user’s return to <a_21><b_19><c_204> signifies their commitment to solidifying their gaming setup, thoughtfully balancing technical enhancements with the enjoyment of gameplay. After establishing their gaming environment, they transitioned back to an equally engaging RPG with <a_92><b_90><c_145>, highlighting their intrinsic draw towards rich narrative experiences that allow for deep player engagement. This seamless shift reveals a thoughtful consideration of both gameplay mechanics and the overall story arc, suggesting that the user values a harmonious blend of technical quality and immersive storytelling. Given this dynamic interaction history, it seems evident that the user will continue to seek out richly woven narrative experiences that challenge the mind and engage the heart. Therefore, it’s only natural that they would be inclined toward <a_92><b_90><c_145> as their predicted next choice, as it perfectly aligns with their penchant for captivating storytelling and interactive gameplay. Reasoning over Semantic IDs Enhances Generative RecommendationConference’17, July 2017, Washington, DC, USA Data Generation For Item-centric Semantic Enrichment Prompt Template for Stage 1: Comprehensive Analysis You are an expert recommendation system analyst analyzing user behavior patterns. Your goal is to reason through the user’s history and predict what the item the user would be interested in, explaining your reasoning process from your analytical perspective in first person. Based on the following product information, generate a comprehensive analysis: Product Title, Brand, Category, Description, Features Please provide: 1. A detailed 2-3 sentence description 2. 2-3 main use cases 3. Target audience 4. 3-5 key features summary 5. 5-8 related keywords Prompt Template For Stage 2: Integration You are a senior copywriter preparing an in-depth narrative for a product dossier. - Source Meta Information: Title, Brand, Categories, Description:, Features - Product Semantic Identifier (use this exact string whenever you mention the product): SID_index - Enrichment from Stage 1: Detailed Description, Target Audience, Key Features, Related Keywords Task: 1. Combine ALL of the information above into a single rich and coherent narrative of at least 10 sentences. Include every important fact,Detailed Description,scenario,Target Audience,audience insight, feature highlight, and keyword context that appears in the sources. 2. Every reference to the product must use the identifier index. Do NOT use the title or any other alias. 3. Ensure the result reads like a rich, flowing paragraph (no bullet points, headings, or enumerations). Maintain a professional and descriptive tone suitable for a product catalog. 4. Highlight how index fits different use cases, why its features matter, and draw from both original data and first-stage enhancements without omitting details. Conference’17, July 2017, Washington, DC, USAHe et al. Data Generation For User-centric Reasoning Enrichment Prompt Template for Stage 1: Reasoning Generation You are an expert recommendation system analyst analyzing user behavior patterns. Your goal is to reason through the user’s history and predict what the item the user would be interested in, explaining your reasoning process from your analytical perspective in first person. Given user interaction history, item descriptions, and reference next item, produce a concise first-person reasoning from an analyst’s perspective to predict what kind of item the user may like in the next interactions. The reference item is only for internal guidance—reason entirely based on interaction history and item descriptions. Never mention or discuss the reference item in your reasoning. Write as a genuine real-time prediction analyzing user behavior patterns. User Interaction history: Chronological item titles Reference next item: Next item title Item Descriptions:Item_descriptions OUTPUT REQUIREMENTS: 1. Output ONLY reasoning monologue in first person (I) as an analyst. Keep concise but detailed.Vary sentence structures to avoid repetition. 2. Analyze general user preferences (genres, themes, attributes, motivations) and engagement patterns based on history. 3. Express potential interests or tendencies rather than deterministic conclusions or single outcomes. 4. Adapt depth to history length: brief key observations for short histories; step-by-step tracing of interest shifts for longer ones. Base predictions on observed patterns. 5. CRITICAL: Always use ONLY the SID format when referring to items. Never use titles, names, or ’Item SID:’ prefixes. 6. Never mention ’reference item’ or imply knowledge of the target. Reason as if predicting blindly. 7. Start directly with reasoning. Do NOT predict a specific next item. End with a non-deterministic summary of likely interests (e.g., ’may enjoy’, ’tends to prefer’). Your Reasoning: Prompt Template For Stage 2: Integration with SIDs Integrate the following information into a single, coherent, natural narrative paragraph: User interaction history: (Chronological item titles & SIDs) Reference next item: (Item Title & SID) Reasoning path from stage 1 OUTPUT REQUIREMENTS: 1. Start your narrative by explicitly reciting the full ’User Interaction history (chronological item SIDs)’ sequence EXACTLY as provided, but use varied and natural opening phrases. Ensure the full sequence is included to establish context. 2. Write in a natural, flowing style—avoid mechanical or formulaic language in the subsequent analysis. Make it read like a genuine narrative. 3. Preserve the essential reasoning insights from the reasoning path—don’t just summarize, but naturally incorporate the key analytical points and logic. 4. When mentioning any item, ALWAYS use its SID (format: <a_X><b_Y><c_Z>)—never use item titles or names. 5. Keep the narrative natural and engaging. Integrated Narrative: