Paper deep dive
From Interests to Semantic IDs: Retrieval-Grounded Credit Assignment for Generative Recommendation
Mengdan Zhu, Yufan Zhao, Yao Zhao, Sophie Di, Tao Di, Yulan Yan, Sridhar Iyer, Liang Zhao
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 92%
Last extracted: 9/26/2026, 2:39:46 AM
Summary
The paper addresses the credit-assignment gap in generative recommendation systems that use Semantic IDs (SIDs) and reasoning traces. Standard training relies on sparse exact-match rewards for the final SID, which fails to credit intermediate reasoning steps. The authors propose a retrieval-grounded credit assignment method where generated interest hypotheses are used as queries for a frozen retriever. Rewards are assigned based on whether these queries retrieve the target item, allowing credit to be localized to specific interest spans rather than the entire rollout. This method improves recommendation performance on Amazon Reviews datasets.
Entities (6)
Relation Signals (5)
Retrieval-Grounded Credit Assignment → addresses → credit-assignment gap
confidence 96% · This creates a credit-assignment gap. We address this gap with retrieval-grounded query attribution.
Retrieval-Grounded Credit Assignment → uses → Frozen Retriever
confidence 95% · A frozen retriever executes every hypothesis as a catalog query, so that each hypothesis becomes independently verifiable rather than judged only through the final SID.
Retrieval-Grounded Credit Assignment → improves → SID recommendation
confidence 94% · Across experiments on three Amazon Reviews datasets, this yields consistent improvements in SID recommendation.
Semantic IDs → enables → autoregressive prediction
confidence 90% · Semantic IDs (SIDs) encode each catalog item as a short token sequence, enabling generative recommenders to predict the next item autoregressively.
SIDReasoner → precedes → Retrieval-Grounded Credit Assignment
confidence 85% · We initialize SID–language alignment and supervised reasoning activation from [6]... and then apply group-relative RL.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Semantic IDs (SIDs) encode each catalog item as a short token sequence, enabling generative recommenders to predict the next item autoregressively. Reasoning-enhanced variants, an increasingly common extension, first generate a textual trace and then decode a next-item SID by beam search. Such recommenders are commonly trained with group-relative policy optimization under an exact-match SID reward, which is sparse in large catalogs. Two failure modes follow. When all rollouts in a group miss the target, the group yields zero advantage and no learning signal. Rollouts sharing the same SID reward receive identical advantages, however much their traces differ. In both cases the reward reflects only the decoded SID, never the reasoning that produced it. This creates a credit-assignment gap. We address this gap with retrieval-grounded query attribution. Each trace is structured into a history summary, a set of interest hypotheses, and a final SID. A frozen retriever executes every hypothesis as a catalog query, so that each hypothesis becomes independently verifiable rather than judged only through the final SID. A rollout is rewarded when any of its queries retrieves the target within the \mbox{top-$K$}, and per-query hit indicators localize that reward to individual hypotheses. Credit is thus assigned at the span level: only hypotheses that individually hit receive positive retrieval advantage, while the retrieval channel never updates the final SID span. Rollouts that share a SID reward can therefore receive different updates. Across experiments on three Amazon Reviews datasets, this yields consistent improvements in SID recommendation. On Video Games, an oracle analysis further reveals the potential of interest-conditioned SID decoding: selecting the target-relevant query among generated interests improves both recall and ranking.
Tags
Links
- Source: https://arxiv.org/abs/2609.29983v1
- Canonical: https://arxiv.org/abs/2609.29983v1
Trouble viewing inline? Open PDF directly →
Full Text
55,431 characters extracted from source content.
Expand or collapse full text
From Interests to Semantic IDs: Retrieval-Grounded Credit Assignment for Generative Recommendation Mengdan Zhu Emory University Atlanta, GA, USA mengdan.zhu@emory.edu Yufan Zhao Microsoft Redmond, WA, USA yufzhao@microsoft.com Yao Zhao Microsoft Redmond, WA, USA yaozhao2@microsoft.com Sophie Di Cornell University Ithaca, NY, USA szd5@cornell.edu Tao Di Microsoft Redmond, WA, USA taodi@microsoft.com Yulan Yan Microsoft Redmond, WA, USA yulanyan@microsoft.com Sridhar Iyer Microsoft Redmond, WA, USA sridhariyer@microsoft.com Liang Zhao Emory University Atlanta, GA, USA liang.zhao@emory.edu Abstract Semantic IDs (SIDs) encode each catalog item as a short token sequence, enabling generative recommenders to predict the next item autoregressively. Reasoning-enhanced variants, an increas- ingly common extension, first generate a textual trace and then decode a next-item SID by beam search. Such recommenders are commonly trained with group-relative policy optimization under an exact-match SID reward, which is sparse in large catalogs. Two failure modes follow. When all rollouts in a group miss the target, the group yields zero advantage and no learning signal. Rollouts sharing the same SID reward receive identical advantages, however much their traces differ. In both cases the reward reflects only the decoded SID, never the reasoning that produced it. This creates a credit-assignment gap. We address this gap with retrieval-grounded query attribution. Each trace is structured into a history summary, a set of inter- est hypotheses, and a final SID. A frozen retriever executes every hypothesis as a catalog query, so that each hypothesis becomes independently verifiable rather than judged only through the fi- nal SID. A rollout is rewarded when any of its queries retrieves the target within thetop-퐾, and per-query hit indicators localize that reward to individual hypotheses. Credit is thus assigned at the span level: only hypotheses that individually hit receive posi- tive retrieval advantage, while the retrieval channel never updates the final SID span. Rollouts that share a SID reward can therefore receive different updates. Across experiments on three Amazon Reviews datasets, this yields consistent improvements in SID rec- ommendation. On Video Games, an oracle analysis further reveals the potential of interest-conditioned SID decoding: selecting the target-relevant query among generated interests improves both recall and ranking. 1 CCS Concepts • Information systems→ Recommender systems. 1 Code available at https://github.com/YuFan-Microsoft/Retrieval-Grounded-Credit- Assignment-for-Generative-Recommendation. Keywords Generative Recommendation, Semantic IDs, Credit Assignment, User Interest Retrieval, Reinforcement Learning 1 Introduction Sequential recommendation predicts the next item a user will inter- act with from their interaction history. Generative recommenders cast this task as autoregressive decoding, which requires an item representation that a language model can emit directly. Semantic ID provides one: each catalog item is represented as a short sequence of discrete tokens obtained by quantizing its content embedding, so that semantically related items share prefixes [15,22]. This rep- resentation allows LLM-based recommenders to generate items within the same autoregressive interface used for language. Recent work first produces a natural-language reasoning trace and then predicts an item [26,34]; in SID-based methods, the trace precedes autoregressive decoding of a next-item SID [6,9,12]. A single roll- out therefore spans two heterogeneous segments: the trace, for which no ground-truth target exists, and a discrete item action, for which the held-out item supplies exact supervision. Once reinforce- ment learning is applied to the joint rollout, the central question is how final recommendation feedback should assign credit to the trace. A common group-relative objective scores each sampled rollout through its generated SID, using exact item correctness or prefix- level matching. These rewards appropriately supervise item decod- ing but provide weak feedback for the intermediate tokens. Such an objective can separate two rollouts only when the reward assigns them different values, and exact match rarely does so in a large catalog: with a finite rollout group, every sampled SID may miss the held-out item [12,30], leaving the reward constant within the group so that normalization produces zero SID-derived advantage. Even when SID rewards do vary, rollouts sharing a reward value receive identical SID-derived credit, and their traces are treated as arXiv:2609.29983v1 [cs.IR] 24 Sep 2026 Zhu et al. equivalent. Prefix-level matching yields a graded score and there- fore mitigates the constant-reward case, but it remains defined over decoded item tokens and still does not evaluate the trace itself. We require the model to state future interests in a fixed format and run each interest as a query against a frozen catalog retriever. Consider a rollout with four interest lines that ends in the wrong SID. If only the second query retrieves the held-out target, per- query evidence records the second line as the only hit. We route positive retrieval advantage to that line, not to the other three lines or the wrong SID. For a non-covering rollout in the same group, no query has positive evidence, so its negative relative advantage is shared across its valid interest lines. The response contains a history summary and several interest queries, followed by a valid SID. A binary rollout reward is one when any query retrieves the held-out target within thetop-퐾; a separate indicator retains the outcome of every query for attribu- tion. Generated queries and personas have previously served as retrieval inputs [1,20]; we use retrieval during training to score the model’s own interest queries. The reward measures target retrieval under the fixed retriever. We initialize SID–language alignment and supervised reasoning activation from [6], adapting the activation trace to our structured interest–SID response, and then apply group- relative RL. During rollout, a prefix trie restricts generated SIDs to valid catalog items. The objective separately normalizes three rewards within each rollout group: a structural reward for format and history citations, a retrieval reward for generated interests, and an exact-match SID reward. The retrieval advantage is then routed with the per-query indicators: hitting interest lines receive the positive signal, while the valid interest lines of a non-covering rollout share its negative signal. It is always zero over the final SID span. If all SID rewards in a group are zero but the retrieval rewards differ, the model can still update the interest lines. If the retrieval rewards are also constant, this additional signal is zero. The generated interests also provide an auxiliary recall interface. We run them against the frozen retriever to form a recall pool. Optionally, the SIDs in this pool define a candidate-specific trie, and the same policy performs beam search with probabilities normalized over that trie. The interests therefore restrict the candidate space, while the SID decoder produces the final recommendation. We summarize our contributions as follows: ●We identify a structural limit of item-level rewards in com- mon sequential recommendation methods: a reward com- puted from the decoded item can become finer only within the SID span, while every token of the preceding trace still receives one rollout-level advantage. ●We therefore propose a process reward on the trace itself. In- terests are emitted in a fixed format and executed as queries against a frozen retriever; per-query outcomes route re- trieval advantage only to the interest spans whose queries retrieve the target, leaving the final SID span under item correctness. The same queries supply candidates at infer- ence. ●Empirically, the proposed method improves SID recom- mendation across three Amazon Reviews datasets. We also quantify how often the retrieval channel reactivates SID- inactive groups with constant exact-match rewards, and use an oracle analysis to demonstrate the potential of interest- conditioned SID decoding when the target-relevant query is selected from the generated interests. 2 Related Work 2.1 Semantic-ID Generative Recommendation Generative recommendation predicts item identifiers rather than scoring the full catalog. TIGER [15] represents each item by a short sequence of codes obtained by quantizing its content embedding, and generates this Semantic ID autoregressively. Later work im- proves how these identifiers are learned or augments them with collaborative signals. VQ-Rec [10] learns vector-quantized repre- sentations for transferable recommendation; LETTER [22] trains the tokenizer with recommendation supervision. EAGER [23] mod- els semantic and behavioral signals in separate streams. LC-Rec [32] and EAGER-LLM [8] further align item tokens with pretrained language models. SIDs have also been used beyond standard next-item predic- tion. Penha et al. [14] study a shared SID space for search and recommendation, while SIGMA [27] uses hybrid item tokens for several instruction-conditioned tasks. Zeng et al. [28] mine latent interests with multimodal LLMs and use them in SID construction and reward design. Quantization, however, does not preserve every property of the original item embedding: Wang et al. [21] report sub- stantial changes in fine-grained neighborhoods after discretization. We retain the tokenizer and SID space of the base recommender, and instead focus on retrieval supervision during reinforcement learning optimization. 2.2Reasoning for Generative Recommendation Recent work adds either explicit or latent reasoning to recommen- dation. R 2 ec [26] uses separate heads for language and recommen- dation, and optimizes them jointly with reinforcement learning. OneRec-Think [12] grounds item tokens in text, activates in-text rea- soning, and then refines the model with a recommendation reward. GREAM [9] combines semantic alignment, a reasoning curriculum, and group policy optimization. SIDReasoner [6] strengthens SID– language alignment before optimizing reasoning trajectories with the final recommendation outcome. OneReason [19] organizes rea- soning around item perception and user-interest cognition, while TwiSTAR [2] learns when slower reasoning is worth its inference cost. Other methods perform the extra computation in latent space. S 2 GR [4] inserts a supervised thinking token before each SID token. PauseRec [5] instead uses learned pause tokens, motivated by the observation that free-form rationales can be sensitive to rationale quality and to the mismatch between text and SID embeddings. These studies also expose a limitation of explicit reasoning: read- able text alone does not establish that the trace is useful. Existing methods usually judge a trace through its final recommendation; we additionally test whether its stated interests can retrieve the target item. 2.3 Process Supervision An exact-match SID reward is sparse because an error in an early SID token can move all sampled trajectories to the wrong catalog From Interests to Semantic IDs: Retrieval-Grounded Credit Assignment for Generative Recommendation branch. PROMISE [3] trains a process reward model over interme- diate SID paths and uses it to guide beam search. HCGRec [30] addresses all-zero rollout groups by supplying a minimal target- prefix hint and assigning different credit to the hinted prefix and sampled suffix. SAPO [33] pairs each reasoning block with one SID token and computes step-aligned match advantages to localize SID- token errors. These methods provide finer supervision along the SID generation path. Our method instead verifies each generated language interest through catalog retrieval and uses per-query hit evidence to route the auxiliary advantage to the corresponding hitting interest line rather than the complete interest block. 2.4 Language-Based Retrieval Another line of work uses generated language as the retrieval in- put. GLoSS [1] produces textual queries from interaction histories and retrieves items with dense semantic search. Wang et al. [20] generate multiple interest personas that serve as an online retrieval source in a large-scale video platform. Such systems use language as the recall interface, whereas SID reasoners directly generate item tokens. Our method connects these two interfaces in one generative policy. The model generates future-interest statements grounded in the interaction history, followed by a catalog-constrained SID. We reward both the structure of the trace and whether its interest statements retrieve the held-out item. At inference time, the same policy can provide queries for candidate recall or directly decode the final SID. 3 Methodology Training uses two pre-RL stages adapted from SIDReasoner [6], followed by retrieval-grounded policy optimization. Stage 1 aligns newly added SID tokens with item language and sequential behav- ior. Stage 2 teaches the model to produce a structured reasoning trace with a history summary and future-interest statements before generating the final SID. Stage 3 samples these trajectories and ap- plies group-relative RL. Each domain uses one fixed SID tokenizer across all stages. Each checkpoint initializes the next stage. Stage 3 supplies feedback at three resolutions. Exact SID match scores the final recommendation, a rule-based trace reward verifies the structure of the reasoning trace and its citations to the inter- action history, and a frozen dense retriever evaluates whether the generated interests retrieve the target from the item catalog. We normalize these rewards separately and apply their advantages to the full response, the reasoning trace, and individual interest lines, respectively. For the retrieval channel, query-attributed span rout- ing sends positive advantage only to interest lines whose queries hit the target. At inference time, the generated interests form an auxiliary recall pool that can optionally define the trie for candidate- constrained SID decoding. 3.1 Problem Definition Let퐼be the item catalog, and let푠푣 = 푠 1 푣,...,푠 퐿 푣denote the fixed퐿-token SID of푣 ∈ 퐼. For a user history퐻 푢 = 푣 1 ,...,푣 푇 , the input푥 푢 consists of a fixed instruction followed by the concatenated history SIDs푠푣 1 ,...,푠푣 푇 ;푠 ∗ 푢 denotes the SID of the held-out next item. Following SIDReasoner [6], we reuse the domain-specific Table 1: Stage-1 SID alignment streams. Training streamPurpose and loss Title↔ SID translationBidirectional grounding (SFT) SID history→ next SIDSID-space recommendation (SFT) SID history→ next-item titleSID-to-title recommendation (SFT) Title history→ next-item titleTitle-space recommendation (SFT) Title history→ next SIDTitle-to-SID recommendation (SFT) Item-level SID–text interleavingItem grounding (full LM) Sequence-level SID–text interleaving Sequence grounding (full LM) General reasoningGeneral-ability retention (SFT) SID tokenizers and item-to-SID mappings produced by its three- level RQ-VAE. Accordingly,퐿 =3, and each of the three codebooks contains 256 entries. The codebooks and item-to-SID mappings remain fixed throughout all training stages. The SID tokens are added to the language model vocabulary alongside the existing text tokens, allowing the same autoregressive policy to generate both the reasoning trace and the final SID. During Stage 3, the policy samples structured trajectories. The 푖-th trajectory contains a reasoning trace with a history summary ℎ 푖 and a future-interest block 푓 푖 , followed by a predicted SID ˆ 푠 푖 : 푦 푖 = (︀ℎ 푖 ; 푓 푖 ; ˆ 푠 푖 ⌋︀, 푦 푖 ∼ 휋 휃 ⋅ ⋃︀ 푥 푢 .(1) Let푞 푖,1 ,...,푞 푖,푀 푖 be the queries in푓 푖 , and let푅 퐾 푞be the top-퐾 items returned by a fixed catalog retriever. We assess each trajec- tory along three dimensions. Its reasoning trace is valid when it satisfies the prescribed format and history-citation rules. The SID path succeeds when ˆ 푠 푖 = 푠 ∗ 푢 . The interest path succeeds when at least one retrieved item has SID 푠 ∗ 푢 . 3.2 Pre-RL Initialization We reuse SIDReasoner’s Stage-1 alignment data [6]. We then con- struct new GPT-4o supervision for Stage 2, adapting the reasoning format to expose retrieval queries. Stage 1: SID Alignment. We train the domain-specific SID tokens using the eight streams in Table 1. The first five rows and general reasoning use chat-formatted, completion-only SFT with loss on the assistant response. The two interleaving streams use full-sequence language modeling to place SIDs in item- and sequence-level text. All streams jointly update the model, including the newly added SID-token embeddings. Stage 2: Structured Reasoning Activation. Each completion target has the following form: <think> <history_summary> - SID(s) => factual observation ... </history_summary> <future_interests> - SID(s) => 푞 푖,1 ... </future_interests> </think> target SID The ellipses denote additional entries in each block. Zhu et al. Figure 1: Grouped trajectory generation, separately standardized reward signals, and token-level credit assignment. Trace fields. A summary line pairs cited history SIDs with a factual observation. A future-interest line states a plausible next interest grounded in its cited history SIDs. For line푗,푞 푖,푗 is the text after=>and is the only part sent to the retriever. The tags locate its line-level token mask in Stage 3. Stage-2 data. Following SIDReasoner, we use an LLM teacher to generate the SFT traces. For each training instance, the teacher receives the interaction history and item metadata together with the held-out next-item metadata, and generates a reasoning trace in the schema above. We pair the trace with the target SID to form the assistant completion. Stage 2 uses completion-only SFT for one epoch, and its checkpoint initializes Stage 3. 3.3 Retrieval-Grounded Policy Optimization Stage 3 assigns credit at two levels. Across the퐺trajectories sam- pled for the same user history, separately normalized SID, trace, and retrieval rewards determine the relative credit of each trajec- tory. Within a trajectory, per-query retrieval evidence attributes retrieval credit to individual interest lines. We combine both levels into token advantages: SID credit applies to the full response, trace credit to the reasoning trace, and retrieval credit only to attributed interest lines. 3.3.1 Catalog-Constrained Rollouts. Each trajectory first generates the two text blocks and the closing</think>marker, then samples its퐿-token SID autoregressively. We build a prefix trie from the catalog SIDs so every sampled sequence completes to a catalog SID. Let푟denote the generated text trace, and suppress the trajectory index in this subsection. At SID position푙, let ˆ 푠 <푙 be the current prefix, let풱 ˆ 푠 <푙 be its valid next tokens, and let푐 푙 = (︀푥 푢 ;푟; ˆ 푠 <푙 ⌋︀be the full autoregressive context. The sampling distribution is 휋 trie 휃 푎 ⋃︀ 푐 푙 = exp푧 휃 푎 ⋃︀ 푐 푙 ∑ 푎 ′ ∈풱 ˆ 푠 <푙 exp푧 휃 푎 ′ ⋃︀ 푐 푙 , 푎 ∈ 풱 ˆ 푠 <푙 .(2) Tokens outside풱 ˆ 푠 <푙 have probability zero. This is the standard language-model softmax restricted to the outgoing edges of the current trie node. 3.3.2 Reward Signals. Each trajectory produces three binary rollout- level rewards. The retrieval channel additionally retains per-query hit indicators for subsequent within-trajectory attribution. SID exact-match reward. The SID reward records exact agree- ment between the sampled and target SIDs: 푟 sid 푖 = 1(︀ ˆ 푠 푖 = 푠 ∗ 푢 ⌋︀. Trace-validity reward. The trace reward is one when the response contains exactly one summary block followed by one interest block and every cited SID occurs in the input history. We write this reward as 푟 trace 푖 = 1(︀valid trace⌋︀. Interest-retrieval reward. A frozen catalog retriever ranks items for each parsed interest query. For a query푞, letrank푠 ⋃︀ 푞be the best rank of any catalog item whose SID is푠, or+∞if no such item is retrieved. For cutoff퐾, we define the query-level evidence and rollout-level retrieval reward as 푧 ret 푖,푗 = 1(︀interest block is valid⌋︀ 1(︀rank푠 ∗ 푢 ⋃︀ 푞 푖,푗 ≤ 퐾⌋︀, 푟 ret 푖 = 1 ⎨ ⎝ ⎝ ⎝ ⎝ ⎪ 푀 푖 ∑ 푗=1 푧 ret 푖,푗 > 0 ⎬ ⎠ ⎠ ⎠ ⎠ ⎮ . (3) Thus,푟 ret 푖 =1 if at least one query retrieves the target within the top-퐾; it is zero when no valid query is parsed. We retain each푧 ret 푖,푗 for line-level routing. We use a top-퐾event rather than the raw retrieval score because the desired behavior is candidate coverage, not score calibration. 3.3.3 Across-Trajectory Credit. The rollout rewards determine which trajectories receive positive or negative relative credit within the group sampled for one user history. For each reward푐 ∈ sid, trace, ret, we compute 휇 푐 푢 = 1 퐺 퐺 ∑ 푖=1 푟 푐 푖 , 휎 푐 푢 = ⟨ ⧸︂ ⧸︂ ⟩ 1 퐺 − 1 퐺 ∑ 푖=1 푟 푐 푖 − 휇 푐 푢 2 , 퐴 푐 푖 = 푟 푐 푖 − 휇 푐 푢 휎 푐 푢 +휖 . (4) From Interests to Semantic IDs: Retrieval-Grounded Credit Assignment for Generative Recommendation Table 2: Reward channels and token-level credit scopes. RewardCriterionCredit scope SIDExact matchFull response TraceValid structureReasoning trace Retrieval Target coverage Attributed interest lines The denominator퐺 −1 matches the sample standard deviation used in the implementation. Following group-relative optimization [16], we standardize rewards within each rollout group. We apply this operation separately to the SID, trace, and retrieval rewards. When- ever a reward varies within a group, its normalized advantages have a comparable scale before weighting. If a reward is constant within a group, it contributes zero advantage. 3.3.4 Token-Level Credit Assignment. The three rewards assign credit to different parts of a trajectory. As the final outcome signal, the SID advantage applies to the full response; the trace advan- tage applies to the reasoning trace; and the retrieval advantage is attributed to interest lines using query-level retrieval evidence. Query-level routing. The trajectory-level advantage퐴 ret 푖 deter- mines the sign and relative magnitude of retrieval credit, while푤 푖,푗 determines which interest lines receive it. For a retrieval-positive trajectory, the binary hit indicators divide the credit equally among the hitting interest lines. For a retrieval-negative trajectory in a mixed group, no query provides positive evidence, so the negative credit is divided equally among all parsed interest lines. We define 푤 푖,푗 = )︀ ⌉︀ ⌉︀ ⌉︀ ⌉︀ ⌉︀ ⌉︀ ⌉︀ ⌉︀ ⌉︀ ⌋︀ ⌉︀ ⌉︀ ⌉︀ ⌉︀ ⌉︀ ⌉︀ ⌉︀ ⌉︀ ⌉︀ ]︀ 푧 ret 푖,푗 ∑ 푀 푖 ℓ=1 푧 ret 푖,ℓ , 푟 ret 푖 = 1, 1 푀 푖 ,푟 ret 푖 = 0, 푀 푖 > 0, 0,푀 푖 = 0. (5) For푀 푖 >0, the line weights sum to one. Positive trajectories select the hitting interest lines, whereas negative trajectories select all parsed interest lines. Token-level composition. For token푡in trajectory푖,푚 line 푖,푗,푡 indi- cates whether it belongs to the푗-th interest line, and푚 trace 푖,푡 indicates whether it belongs to the reasoning trace. These masks localize the retrieval and trace advantages to their corresponding spans. If an interest line cannot be located unambiguously, its mask is zero. The retriever evaluates only the query text, while the routed retrieval advantage is applied to the corresponding interest line. Combining these credit scopes gives the token-level advantage: 퐴 푖,푡 = 퐴 sid 푖 +휆 trace 푚 trace 푖,푡 퐴 trace 푖 +휆 ret 푀 푖 ∑ 푗=1 푤 푖,푗 푚 line 푖,푗,푡 퐴 ret 푖 .(6) The retrieval advantage is divided across the attributed interest lines through푤 푖,푗 and applied uniformly to their tokens. Positions in the final SID span receive only 퐴 sid 푖 . Table 2 summarizes the resulting token-level credit scopes. 3.3.5 SID-Inactive Groups. A rollout group is SID-inactive when its SID exact-match rewards are constant, which gives퐴 sid 푖 =0 for every trajectory. The common all-zero case occurs when none Table 3: Dataset statistics after preprocessing. CategoryUsers Items Train Val. Test Avg. len. Video Games 6,142 3,858 49,133 6,142 6,1426.45 Office4,866 3,459 38,924 4,866 4,8665.97 Industrial4,533 3,686 36,259 4,532 4,5335.96 of the sampled SIDs matches the target. If retrieval rewards vary within such a group, the retrieval channel can still assign nonzero credit to attributed interest lines and distinguish trajectories whose queries retrieve the target from those whose queries do not. 3.3.6 Policy Optimization. Let휌 푖,푡 be the current-to-old probability ratio for response token푡. Both its numerator and denominator use standard language-model probabilities on reasoning tokens and the trie-restricted distribution in Eq.(2)on final SID tokens, thereby matching the rollout action space. Let푑 KL 푖,푡 denote the token-level KL divergence from the frozen pre-RL policy, evaluated over the same action space. With푁 = ∑ 퐺 푖=1 ⋃︀푦 푖 ⋃︀and coefficient훽 KL ≥0, we optimize ℒ actor = − 1 푁 퐺 ∑ 푖=1 ⋃︀푦 푖 ⋃︀ ∑ 푡=1 [︀ℓ PPO 휌 푖,푡 ,퐴 푖,푡 − 훽 KL 푑 KL 푖,푡 ⌉︀.(7) Here ℓ PPO is the dual-clip PPO surrogate [25]. 3.4 SID Decoding with Auxiliary Interest Recall SID decoding. Our primary inference mechanism is SID decoding. Given a user history, the policy first generates a structured trace. Conditioned on the history and this trace, catalog-constrained beam search returns a ranked list퐶 sid of valid SID candidates, which serves as the direct recommendation output. Auxiliary interest recall. The retrieval-supervised interests ad- ditionally provide an auxiliary recall interface. We issue each gen- erated interest query to the frozen catalog retriever used during training, merge the per-query top-퐾results, and remove duplicate SIDs to obtain a recall pool퐶 ret . This pool exposes catalog coverage captured by the generated interests. Optional candidate-constrained SID decoding. The recall pool can also define a candidate-specific SID trie. We build this trie from the SIDs in퐶 ret and, conditioned on the same history and trace, per- form beam search with probabilities normalized over the candidate trie. This returns a ranked list퐶 rcd through candidate-constrained generation rather than reranking candidates by their full-catalog SID probabilities. 4 Experiments 4.1 Experimental Setup Datasets and protocol. We use the Video Games, Office Products, and Industrial and Scientific categories of the Amazon Reviews corpus [13]. Following the processed data used by SIDReasoner [6], we retain interactions from October 2016 through November 2018, apply 5-core filtering, sort each user’s interactions chronolog- ically, and truncate histories to the most recent 10 items. The last interaction of each user is held out for test, the second-to-last for Zhu et al. Table 4: Overall recommendation performance. The best result in each column is shown in bold. The SID-plus-trace control and our method report means over three training seeds; their maximum standard deviations across all columns are 0.0053 and 0.0039, respectively. An asterisk (*) indicates a statistically significant improvement over the SID-plus-trace-validity control under a paired two-sided 푡-test over test users (푝 < 0.05). Video GamesOffice ProductsIndustrial and Scientific MethodRecall@5NDCG@5Recall@10NDCG@10Recall@5NDCG@5Recall@10NDCG@10Recall@5NDCG@5Recall@10NDCG@10 Traditional sequential recommenders GRU4Rec0.03290.02190.05990.03050.06820.04800.09740.05740.07880.05780.10300.0649 SASRec0.05010.03450.07230.04160.10190.08240.11670.08710.08070.06470.09640.0697 Caser0.03760.02410.06590.03320.08800.06630.11140.07380.06640.05280.08520.0588 Generative recommenders without reasoning TIGER0.04890.03000.07630.04020.12700.10370.14290.11210.10030.08230.13250.0924 HSTU0.05390.03960.07460.04620.12040.10690.13230.11070.10080.08980.11380.0940 LETTER0.04450.02940.07090.03780.13150.10740.15200.11390.10800.08500.13890.0950 LCRec0.04410.02740.08760.04120.09640.06990.14870.08670.08050.05200.13300.0687 Reasoning-based recommenders ReaRec0.05680.03810.08430.04700.11730.09880.13850.10570.09730.07960.12050.0870 R 2 ec0.06550.03990.09310.05250.11470.08940.14860.10040.08800.07740.12530.0774 SIDReasoner0.07100.04600.10310.05630.13730.11190.16480.12080.11090.09050.14380.1010 Controlled Stage-3 RL ablations Stage-2 SFT checkpoint (no RL)0.05910.04070.08580.04930.12970.10740.15060.11420.11250.08930.13700.0972 Stage-3 RL: SID exact-match reward0.05830.03990.08610.04890.12810.10460.15050.11190.10980.08560.13710.0944 + Trace-validity reward0.05910.04040.08560.04930.12720.10510.15120.11240.11160.08680.13870.0953 + Query-attributed retrieval reward (Ours)0.0879 * 0.0694 * 0.1195 * 0.0796 * 0.15150.12510.1765 * 0.1332 * 0.12910.10430.1602 * 0.1144 * AbsoluteΔ vs. SID+trace+0.0288+0.0290+0.0339+0.0303+0.0243+0.0200+0.0253+0.0208+0.0175+0.0175+0.0215+0.0191 validation, and earlier interactions form sliding-window training pairs. Table 3 summarizes the resulting data. Evaluation metrics. We use the SIDReasoner evaluator. Table 4 in- cludes its reported baselines [6]. For each test instance, the evaluator performs catalog-constrained beam search using a trie containing all catalog SIDs and returns a ranked beam of ten valid SIDs. Evalu- ation uses no sampled negatives. We report Recall and NDCG at cutoffs 5 and 10. For interest retrieval, Interest Recall@50 (I-R@50) is the fraction of examples for which at least one generated interest retrieves the held-out target within the frozen retriever’s top 50 results. Our model generates three to four interests per instance, with an average of 3.8. Baselines. We compare against three families of recommenders: ●Traditional sequential recommenders: GRU4Rec [7], SASRec [11], and Caser [18], which rank items directly from the interaction sequence. ●Generative recommenders without reasoning: TIGER [15], HSTU [29], LETTER [22], and LCRec [32]. ● Reasoning-based methods: ReaRec [17], R 2 ec [26], and SIDReasoner [6]. Implementation details. Based on Qwen3-1.7B [24], we fine-tune all parameters in every stage, paired with the same domain-specific three-level RQ-VAE tokenizer. Stage 1 uses AdamW for up to five epochs with a learning rate of 2×10 −5 . Stage 2 uses AdamW for one epoch with a learning rate of 1 × 10 −5 . Stage 3 samples 16 trajectories per prompt with temperature 1.0 and top-푝 =1.0. We use a global batch of 256 prompts, maximum prompt and response lengths of 1,024 tokens each, an actor learn- ing rate of 7×10 −7 , and dual-clip PPO with a clip ratio of 0.2 and dual-clip constant 3.0. Training runs for at most 10 epochs, with checkpoint selection by validation NDCG@10. The default configu- ration omits KL regularization because a Video Games sensitivity check found no material effect on recommendation performance. The trace-validity and retrieval rewards are each weighted by 0.1. Each catalog document contains labeled title, optional brand, and description lines. Their Qwen3-Embedding-4B embeddings [31] are precomputed and frozen; each generated interest is encoded online and matched by exact cosine search. We use퐾 =50 by default and report sensitivity across 퐾 ∈ 10, 20, 50, 100. All hyperparameters, checkpoint selection, and reward cutoffs are selected using validation data. The test set is evaluated once after the experimental protocol is frozen. Across all three domains, the SID-plus-trace control and our method report the mean and standard deviation over three Stage-3 seeds to quantify training variability. 4.2 Overall Recommendation Performance RQ1 asks whether retrieval-grounded credit improves direct full- catalog SID recommendation. Table 4 compares Stage-2 SFT with three incremental RL variants. The matched no-retrieval control is the SID-exact-match-plus-trace-validity variant. All three RL variants start from the same Stage-2 checkpoint and use the same data order, training budget, SID action space, and KL configuration. Retrieval-grounded credit consistently improves direct SID rec- ommendation. Relative to the matched no-retrieval control, our method improves Recall@10 by 15.5%–39.6% and NDCG@10 by 18.5%–61.5% across the three domains. The gains are largest on Video Games and more moderate on Office Products and Industrial and Scientific. R 2 ec shows a similar cross-domain pattern, suggest- ing that the benefit of explicit reasoning may depend on how well item metadata aligns with semantic knowledge available to the language model. Improvements also hold at cutoff 5. In contrast, SID exact-match and trace-validity optimization remain close to the Stage-2 checkpoint. This pattern is consistent with the gain not being explained by RL or structural supervision alone. Section 4.3 examines the additional learning signal supplied by retrieval. Across all baselines in Table 4, our method achieves the best result on all 12 metrics. From Interests to Semantic IDs: Retrieval-Grounded Credit Assignment for Generative Recommendation 020406080100 40 50 60 70 80 Groups without target credit (%) (a) SID credit only 020406080100 40 50 60 70 80 (b) SID + retrieval credit 020406080100 0 3 6 9 12 Δ Interest Recall@50 (p) (c) Reward-retriever coverage Stage-3 training progress (%) Video GamesOffice ProductsIndustrial & Scientific Figure 2: Stage-3 no-signal and retrieval-reactivation dynamics. (a) SID no-signal rate: the percentage of groups with constant binary SID exact-match rewards. (b) Joint no-signal rate: the percentage with both constant SID and constant retrieval rewards. (c) Validation Interest Recall@50 difference, in percentage points, between our method and the matched SID-plus-trace control under the reward retriever. 4.3 Retrieval Credit and Routing RQ2 examines two questions: how often retrieval feedback sup- plies a nonzero advantage when the SID reward is constant, and whether assigning this advantage to queries that produce a re- trieval hit improves recommendation over coarser credit scopes. We analyze the SID and retrieval channels because both provide target-dependent feedback. The trace-validity reward is nearly al- ways constant within a group and therefore usually produces zero group-relative advantage. Let풢 푡 be the rollout groups in training window푡, with퐺trajecto- ries in each group. A reward channel produces zero group-relative advantage when its rewards are constant across the group. For 푐 ∈ sid, ret, let푢 푐 푔 indicate this no-signal condition for channel푐 in group 푔. The no-signal and conditional reactivation rates are 푢 푐 푔 = 1]︀max 푖 푟 푐 푔,푖 = min 푖 푟 푐 푔,푖 ︀, 풵 sid 푡 = 1 ⋃︀풢 푡 ⋃︀ ∑ 푔∈풢 푡 푢 sid 푔 , 풵 sid+ret 푡 = 1 ⋃︀풢 푡 ⋃︀ ∑ 푔∈풢 푡 푢 sid 푔 푢 ret 푔 , ℛ react 푡 = 1 − 풵 sid+ret 푡 풵 sid 푡 , 풵 sid 푡 > 0. (8) 풵 sid is the fraction of groups with no SID-derived advantage, while 풵 sid+ret is the fraction for which both SID and retrieval produce zero group-relative advantage. Their relative reduction,ℛ react , is the fraction of SID-inactive groups reactivated by retrieval. Because both rewards are binary, a no-signal group can be all zero or all one; both cases produce zero group-relative advantage. The reported rates therefore characterize optimization signal, not exact-match sparsity alone. Figure 2 plots these quantities over training. Panels (a) and (b) aggregate all logged groups from the same run of our method. Each point uses a 5%-wide window centered at the displayed progress, computed as completed Stage-3 updates divided by total updates. Panel (c) evaluates the corresponding checkpoints on a fixed valida- tion set and reports the Interest Recall@50 difference, in percentage points (p), between our method and the matched SID-plus-trace control. Training dynamics. The SID channel produces zero group-relative advantage in 69.1%–80.8% of groups across the displayed domains and training windows. By construction,풵 sid+ret ≤ 풵 sid ; empirically, the gap never falls below 27.8 p on Games, 22.5 p on Office, or 14.7 p on Industrial. In the last displayed training window,ℛ react is 37.8% on Games, 37.7% on Office, and 19.6% on Industrial. Panel (c) evaluates the generated interests with the same re- triever used to compute the training reward. At the final displayed checkpoint, our method improves Interest Recall@50 over the SID- plus-trace control by 11.0 p on Video Games, 2.2 p on Office Products, and 3.6 p on Industrial and Scientific. The retrieval reward has its largest measured effect on Video Games. Because training and evaluation use the same retriever, this panel measures learning progress rather than independent generalization. Credit-routing ablations. Table 5 compares five Stage-3 variants in all three domains. Each result averages three seeds. Within a domain, the variants share the Stage-2 checkpoint and training configuration, and all retain the SID exact-match and trace-validity rewards. The SID-prefix variant adds the partial reward used by SIDReasoner [6]. Let푚 푖 be the length of the longest correct prefix between the predicted and target SIDs. The reward is 푟 prefix 푖 = 1 2 퐿−푚 푖 ,(9) which increases toward one as more leading SID tokens match and equals one for an exact match. After within-group normalization, its advantage is applied to the full response, matching the scope of the exact-match SID advantage. Zhu et al. 0.6B4B (Ours) 8B 0 10 20 30 Metric value (%) 30.53 31.49 31.31 10.83 11.95 11.87 7.12 7.96 7.74 (a) Reward encoder capacity (K train = 50) 102050 (Ours) 100 29.99 31.29 31.49 31.62 11.72 11.85 11.95 11.84 7.81 7.90 7.96 7.82 (b) Training cutoff (4B encoder) I-R@50Recall@10NDCG@10 Figure 3: Reward-design sensitivity on Video Games. Bars report absolute Interest Recall@50, Recall@10, and NDCG@10 in percent. Panel (a) varies reward-encoder capacity at a fixed퐾 train =50; panel (b) varies the training cutoff with the 4B encoder. Shaded groups mark our selected 4B encoder and 퐾 train = 50 cutoff. Table 5: Credit-routing ablations across three domains, aver- aged over three seeds. GamesOfficeIndustrial VariantR@10N@10R@10N@10R@10N@10 SID exact + trace0.08560.04930.15120.11240.13870.0953 SID-prefix0.09960.05690.15870.12090.14980.1001 Full broadcast0.09530.05590.16440.11710.14420.1039 Interest block0.10730.06100.16620.12560.15410.1082 Hit query (Ours)0.11950.07960.17650.13320.16020.1144 The remaining variants use the same retrieval reward and differ only in where its advantage is applied. Broadcast applies it to the full response, while block routing confines it to the future-interest block. Hit-aware routing further uses the indicators in Eq.(5): on a positive rollout, only queries that retrieve the target receive retrieval credit. These controls separate the effect of denser SID supervision from the effects of retrieval feedback and query-level attribution. All four auxiliary variants improve on the SID-plus-trace control for every domain and metric. Routing retrieval advantage to the interest block consistently beats broadcasting it to the full response. Hit-query routing performs best on all six metrics. Compared with block routing, its Recall@10 and NDCG@10 gains are 11.4% and 30.5% on Video Games, 6.2% and 6.1% on Office Products, and 4.0% and 5.7% on Industrial and Scientific. This consistent ordering sup- ports assigning retrieval credit to the queries responsible for target coverage. 4.4 Reward Validity and Interest Utility Interest–SID alignment. Table 6 cross-tabulates Interest Recall@50 and SID Recall@10 on Video Games. SID Recall@10 is 27.56% when the interests retrieve the target and 4.78% otherwise. This 22.78 percentage points gap has a 95% bootstrap CI of [20.70, 24.88]. The odds ratio is 7.58 (95% bootstrap CI [6.40, 9.06]). The overlap is partial. Among targets absent from the SID beam, 25.91% are still retrieved by a generated interest. Retrieval success Table 6: Overlap between interest retrieval at 50 and direct SID decoding at 10 for our model on Video Games. Percent- ages are over all 6,142 test instances. SID hit@10SID miss@10Total Interest hit@50533 (8.68%)1,401 (22.81%)1,934 (31.49%) Interest miss@50201 (3.27%)4,007 (65.24%)4,208 (68.51%) Total734 (11.95%)5,408 (88.05%)6,142 thus tracks SID accuracy while exposing target coverage that the final SID outcome does not capture. Potential of interest conditioned SID decoding. Figure 5 compares direct SID decoding with candidate-constrained variants on Video Games. A trie stores candidate SIDs by prefix. At each decoding step, the model can choose only a token that continues one of these candidate SIDs. The random variant selects one generated interest and builds a trie from its top-50 candidates. The all-interest variant merges and deduplicates the top-50 candidates from every generated interest. The oracle uses the held-out target rank to choose the best interest already produced by the model. We evaluate its cosine-ranked top 10 directly, then use its top-50 candidates to constrain SID beam search. A random interest performs substantially worse than global SID decoding. Merging all interests recovers this loss and slightly exceeds global SID decoding on both metrics, showing that the generated interests provide complementary candidate coverage de- spite differing in individual utility. Oracle retrieval improves recall, while SID decoding over the oracle trie gives the strongest recall and ranking. The random and oracle variants use the same pool size, so their gap comes from which interest is selected. This upper bound shows the potential of interest conditioned SID decoding. Se- lecting an interest without the target remains a promising direction for future work. From Interests to Semantic IDs: Retrieval-Grounded Credit Assignment for Generative Recommendation USER HISTORY Original interaction titles REASONING BLOCK <history_summary> <future_interests>Catalog retrieval PREDICTION Target item Target SID Direct SID decoding 1 Watch Dogs 2 - PlayStation 4 2 Ratchet & Clank - PlayStation 4 3 Horizon Zero Dawn - PlayStation 4 4 CorpCo 6ft AC Power Cord for Sony PlayStation 4 (PS4) 5 Deadpool - PlayStation 4 6 DualShock 4 Wireless Controller for PlayStation 4 - Jet Black [Old Model] Uncharted 4: A Thief's End - PlayStation 4 <a_201><b_31><c_107> •Four PlayStation 4 software entries are described as single-player action-adventure titles with campaign-focused play, exploration, and combat. •Two PlayStation 4 hardware entries cover console power connectivity and a wireless controller. •The user may be interested in additional PlayStation 4 single-player action-adventure titles, with the history bridge being repeated campaign-driven exploration, combat, and story-led progression. •The user may be interested in PlayStation 4 setup and play-session hardware, with the history bridge being one entry focused on console power and another on wireless controller use. •The user may be interested in PlayStation 4 sandbox-style hacking or systemic-approach experiences, with the explicit history bridge being open-world mission structure built around manipulating connected infrastructure and technology. •The user may be interested in PlayStation 4 character-led superhero action with irreverent humor, with the explicit history bridge being a comic-book lead, fast-paced combat, and comedic tone. Rank 1 ✓ Rank >100 ✗ Rank 80 ✗ Rank >100 ✗ SID beam rank: 2 generatedecode Figure 4: A Video Games case where both paths find the target. The action-adventure interest retrieves it at rank 1, while the other interests miss at the training cutoff. The SID beam ranks it second. History SIDs are shown as item titles, line-level citations are omitted, and bold text highlights each statement’s key phrase. 0.02.55.07.510.012.515.0 Metric value (%) Global SID decoding Random interest + trie All-interest union + trie Oracle interest retrieval Oracle interest + trie 11.95 8.12 12.51 13.20 14.93 7.96 5.78 8.11 7.93 9.33 Recall@10NDCG@10 Figure 5: Interest-conditioned SID decoding on Video Games. Each interest retrieves its top-50 candidates. The all-interest variant merges and deduplicates candidates from all gener- ated queries. The oracle uses the target rank to select the best generated query and therefore represents an upper bound. 4.5 Sensitivity to Reward Encoder and Cutoff Figure 3 varies the Qwen3-Embedding encoder size at퐾 train = 50 and the training cutoff with the 4B encoder. All runs share the same Stage-2 initialization and Stage-3 configuration, and all are evaluated at I-R@50. The 4B encoder improves over 0.6B and matches 8B while using half the online encoder capacity. This suggests that reward quality saturates before the largest encoder. Recommendation accuracy changes little across cutoffs. Interest Recall increases with퐾 train , while the SID metrics are highest at 50 in this sweep. The divergence shows that maximizing retrieval coverage does not necessarily provide the best signal for SID learn- ing. A larger cutoff accepts more distant matches under the same binary reward, whereas a smaller cutoff is more selective. We use 퐾 train = 50 as the middle ground and retain the 4B encoder. 4.6 Qualitative Case Study Figure 4 shows one Video Games test case in which both paths find the target. Only the action-adventure interest retrieves Uncharted 4: A Thief’s End, at rank 1. The other interests miss, while the direct SID beam ranks the target second. 5 Conclusion We studied credit assignment in reasoning-enhanced SID recom- mendation, where a common issue is that the reward computed from the decoded item can be refined only within the SID span: it cannot separate rollouts that receive the same reward, and it carries no information when that reward is constant within a group. We proposed retrieval-grounded query attribution, which executes the generated interests against a frozen catalog retriever and routes re- trieval advantage only to the interest spans that retrieve the target, leaving the final SID positions under item correctness. Across three Amazon Reviews datasets, the proposed method im- proves SID recommendation over a matched SID-and-trace control and restores group-relative learning signals in many SID-inactive Zhu et al. groups. Ablations show that hit-aware routing outperforms coarser credit scopes, and the generated queries complement direct SID beam search. More broadly, these results show that intermediate predictions, once made executable, can supply localized supervision without a separate process reward model. 6 Ethical Considerations Our experiments use the public Amazon Reviews benchmark and do not collect new user data. Nevertheless, methods that infer future interests from interaction histories may expose sensitive prefer- ences if deployed on identifiable logs. Generated interests may also reproduce popularity and representation biases in catalog metadata and the retriever, which can narrow user exposure and reinforce filter bubbles. The generated trace is optimized for retrieval utility, not verified as a faithful explanation of model behavior, and should not be presented to users without further validation. A deployment should therefore restrict access to histories, avoid storing generated queries, and audit retrieval rewards across user and item groups. References [1] Krishna Acharya, Aleksandr V Petrov, and Juba Ziani. 2025. Gloss: Generative language models with semantic search for sequential recommendation. arXiv preprint arXiv:2506.01910 (2025). [2] Shiteng Cao, Kaian Jiang, Yunlong Gong, and Zhiheng Li. 2026. TwiSTAR: Think Fast, Think Slow, Then Act, Generative Recommendation with Adaptive Reasoning. arXiv preprint arXiv:2605.11553 (2026). [3] Chengcheng Guo, Kuo Cai, Yu Zhou, Qiang Luo, Ruiming Tang, Han Li, Kun Gai, and Guorui Zhou. 2026. PROMISE: Process Reward Models Unlock Test-Time Scaling Laws in Generative Recommendations. arXiv preprint arXiv:2601.04674 (2026). [4] Zihao Guo, Jian Wang, Ruxin Zhou, Youhua Liu, Jiawei Guo, Jun Zhao, Xiaoxiao Xu, Yongqi Liu, and Kaiqiao Zhan. 2026. S2GR: Stepwise Semantic-Guided Reasoning in Latent Space for Generative Recommendation. In Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2. 1498–1507. [5] Yinhan He, Liam Collins, Bhuvesh Kumar, Jundong Li, Neil Shah, and Donald Loveland. 2026. Implicit Reasoning for Large Language Model-based Generative Recommendation. arXiv preprint arXiv:2606.14142 (2026). [6]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 Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2. 1638–1649. [7] Balázs Hidasi, Alexandros Karatzoglou, Linas Baltrunas, and Domonkos Tikk. 2015. Session-based recommendations with recurrent neural networks. arXiv preprint arXiv:1511.06939 (2015). [8] Minjie Hong, Yan Xia, Zehan Wang, Jieming Zhu, Ye Wang, Sihang Cai, Xiaoda 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. [9]Minjie Hong, Zetong Zhou, Zirun Guo, Ziang Zhang, Ruofan Hu, Weinan Gan, Jieming Zhu, and Zhou Zhao. 2025. Generative reasoning recommendation via llms. arXiv preprint arXiv:2510.20815 (2025). [10]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. [11]Wang-Cheng Kang and Julian McAuley. 2018. Self-attentive sequential recom- mendation. In 2018 IEEE international conference on data mining (ICDM). IEEE, 197–206. [12]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). [13]Jianmo Ni, Jiacheng Li, and Julian McAuley. 2019. Justifying recommendations using distantly-labeled reviews and fine-grained aspects. In Proceedings of the 2019 conference on empirical methods in natural language processing and the 9th international joint conference on natural language processing (EMNLP-IJCNLP). 188–197. [14]Gustavo Penha, Edoardo D’Amico, Marco De Nadai, Enrico Palumbo, Alexandre Tamborrino, Ali Vardasbi, Max Lefarov, Shawn Lin, Timothy Heath, Francesco Fabbri, et al.2025. Semantic ids for joint generative search and recommendation. In Proceedings of the Nineteenth ACM Conference on Recommender Systems. 1296– 1301. [15] 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. [16]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). [17]Jiakai Tang, Sunhao Dai, Teng Shi, Jun Xu, Xu Chen, Wen Chen, Jian Wu, and Yuning Jiang. 2026. Think before recommend: Unleashing the latent reasoning power for sequential recommendation. IEEE Transactions on Knowledge and Data Engineering (2026). [18]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. [19]OneRec Team, Biao Yang, Boyang Ding, Chenglong Chu, Dunju Zang, Fei Pan, Han Li, Hao Jiang, Honghui Bao, Huanjie Wang, et al.2026. OneReason Technical Report. arXiv preprint arXiv:2606.06260 (2026). [20] Haoting Wang, Haokai Lu, Zheyun Feng, Jenny Huang, Yifat Amir, Gregory Hink- son, Ben Most, Zelong Zhao, Yixin Kelly Cui, Rein Zhang, et al.2026. LLM-Based User Personas for Recommendations at Scale. arXiv preprint arXiv:2606.12198 (2026). [21] Junting Wang, Xinrui He, Yunzhe Li, and Hari Sundaram. 2026. Understand- ing Semantic IDs: From Item Representation to Item Selection in Generative Recommendation. arXiv preprint arXiv:2607.24995 (2026). [22]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. [23]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. [24]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). [25]Deheng Ye, Zhao Liu, Mingfei Sun, Bei Shi, Peilin Zhao, Hao Wu, Hongsheng Yu, Shaojie Yang, Xipeng Wu, Qingwei Guo, et al.2020. Mastering complex control in moba games with deep reinforcement learning. In Proceedings of the AAAI conference on artificial intelligence, Vol. 34. 6672–6679. [26] Runyang You, Yongqi Li, Xinyu Lin, Xin Zhang, Wenjie Wang, Wenjie Li, and Liqiang Nie. 2026. R 2 ec: Towards Large Recommender Models with Reasoning. Advances in Neural Information Processing Systems 38 (2026), 62376–62405. [27] Yang Yu, Lei Kou, Huaikuan Yi, Bin Chen, Yayu Cao, Lei Shen, Chao Zhang, Bing Wang, and Xiaoyi Zeng. 2026. SIGMA: A Semantic-Grounded Instruction- Driven Generative Multi-Task Recommender at AliExpress. arXiv preprint arXiv:2602.22913 (2026). [28]Yangchen Zeng, Zhenyu Yu, Zhiyuan Hu, Wenxin Zhang, Jinze Wang, and Rongfeng Guo. 2026. DeepInterestGR: Mining Deep Multi-Interest Using Multi- Modal LLMs for Generative Recommendation. arXiv preprint arXiv:2602.18907 (2026). [29]Jiaqi Zhai, Lucy Liao, Xing Liu, Yueming Wang, Rui Li, Xuan Cao, Leon Gao, Zhaojie Gong, Fangda Gu, Michael He, et al.2024. Actions speak louder than words: Trillion-parameter sequential transducers for generative recommenda- tions. arXiv preprint arXiv:2402.17152 (2024). [30]Kangning Zhang, Haotian Fang, Xukun Luo, Hao Yin, Yang Gao, Peng Yan, Weiwen Liu, Weinan Zhang, and Yong Yu. 2026. HCGRec: Hint-Conditioned Generative Recommendation with Semantic IDs. arXiv preprint arXiv:2608.11980 (2026). [31]Yanzhao Zhang, Mingxin Li, Dingkun Long, Xin Zhang, Huan Lin, Baosong Yang, Pengjun Xie, An Yang, Dayiheng Liu, Junyang Lin, et al.2025. Qwen3 embedding: Advancing text embedding and reranking through foundation models. arXiv preprint arXiv:2506.05176 (2025). [32]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. [33]Zaiyi Zheng, Liang Wu, Guanghui Min, Yaochen Zhu, Liangjie Hong, Chen Chen, and Jundong Li. 2026. SAPO: Step-Aligned Policy Optimization for Reasoning- Based Generative Recommendation. arXiv preprint arXiv:2605.17648 (2026). [34]Mengdan Zhu, Yufan Zhao, Tao Di, Yulan Yan, and Liang Zhao. 2026. Learning User Interests via Reasoning and Distillation for Cross-Domain News Recom- mendation. arXiv preprint arXiv:2602.15005 (2026).