S-GRec: Semantic-Aware Generative Rec

• S-GRec is a generative recommendation framework that decouples a lightweight online generator from an offline LLM-based semantic judge to derive stable semantic rewards.
• It employs a two-stage Personalized Semantic Judge (PSJ) to provide interpretable aspect evidence through supervised discrete scoring of profile, novelty, and context.
• The Asymmetric Advantage Policy Optimization (A2PO) fuses business and semantic advantages, ensuring semantic rewards boost rather than overpower core business objectives.

S-GRec denotes “Personalized Semantic-Aware Generative Recommendation with Asymmetric Advantage”, a semantic-aware generative recommendation framework that decouples an online lightweight generator from an offline LLM-based semantic judge for train-time supervision (Jiang et al., 11 Feb 2026). It was introduced to address a specific limitation of generative recommendation trained from behavioral logs: such logs provide weak supervision on underlying user intent, even when the generator directly models item generation as a sequence of semantic IDs (SIDs). S-GRec addresses this limitation through a two-stage Personalized Semantic Judge (PSJ) that produces interpretable aspect evidence and stable semantic rewards, and through Asymmetric Advantage Policy Optimization (A2PO), which anchors optimization on business rewards while injecting semantic advantages only when they are consistent (Jiang et al., 11 Feb 2026).

1. Problem Formulation and Motivation

S-GRec operates in the setting of generative recommendation, where the model directly generates a sequence of SIDs for the target item. Given unified user input $x$ comprising profile, behavior history, and context, the policy is defined as

$$\pi_\theta(y\mid x)=\prod_{t=1}^{T}\pi_\theta(y_t\mid x,y_{<t}).$$

The motivating claim is that behavioral logs mainly teach the model correlations rather than user intent, whereas LLMs contain semantic priors that could provide richer supervision. The paper identifies two obstacles to direct LLM adoption in industrial recommendation: cost-scalability conflict and objective-alignment conflict. LLM inference is described as too slow, memory-heavy, and expensive for millisecond-latency, high-throughput serving, while semantic judgments can conflict with platform objectives such as CTR, eCPM, CVR, and GMV (Jiang et al., 11 Feb 2026).

This problem framing distinguishes S-GRec from approaches that simply fuse semantic and business rewards. The paper explicitly states that simple reward-level fusion is brittle because the two rewards differ in scale and distribution. The design objective is therefore not merely to add semantic information, but to do so in a way that preserves business alignment and eliminates any requirement for real-time LLM inference at serving time (Jiang et al., 11 Feb 2026).

2. Personalized Semantic Judge

The Personalized Semantic Judge (PSJ) is the semantic reward provider in S-GRec. It has a two-stage design consisting of aspect-level semantic scoring and user-conditional preference aggregation. For a user-input and item pair $(x,a)$, PSJ produces a four-dimensional aspect vector $\mathbf{s}(x,a)\in\mathbb{R}^4$. The four aspects are User Profile Relevance, Future Interest Relevance, Novelty, and Contextual Relevance (Jiang et al., 11 Feb 2026).

The appendix-defined scoring rules are discrete and interpretable. Profile relevance uses levels like $-1,-0.5,0,0.5,1$; future interest and novelty use $[0,0.5,1]$; contextual relevance uses $[-1,0]$. Point-wise samples are constructed from model-generated candidate sets and online exposure logs, including clicked, converted, and not-clicked exposures. An LLM generates initial labels, and human annotators verify and correct them. For each semantic dimension, the reward combines exact-match accuracy and order-consistency, and the aspect scorer is trained with supervised fine-tuning and then further aligned with GRPO plus KL regularization (Jiang et al., 11 Feb 2026).

Aspect scores are not used directly as the final reward. PSJ learns a user-dependent aggregation function

$$\mathbf{w}(x)=g_\psi(x)\in\mathbb{R}^4,\qquad w_d(x)\ge 0,\quad \sum_{d=1}^{4} w_d(x)=1,$$

which produces the holistic semantic score

$$s_{\mathrm{hol}}(x,a)=\mathbf{w}(x)^\top \mathbf{s}(x,a)=\sum_{d=1}^{4} w_d(x)\, s_d(x,a).$$

The aggregation stage is trained from pairwise preference data of the form $(x,a_i,a_j,y_{ij})$, where $$\pi_\theta(y\mid x)=\prod_{t=1}^{T}\pi_\theta(y_t\mid x,y_{<t}).$$0 indicates whether $$\pi_\theta(y\mid x)=\prod_{t=1}^{T}\pi_\theta(y_t\mid x,y_{<t}).$$1 or $$\pi_\theta(y\mid x)=\prod_{t=1}^{T}\pi_\theta(y_t\mid x,y_{<t}).$$2. The paper argues that pairwise comparison is more stable than absolute holistic scoring, because annotators can more reliably decide which of two candidates is better and such comparisons expose trade-offs between aspects. The final output of PSJ is therefore both interpretable aspect evidence and a holistic semantic score used as reward for recommendation optimization (Jiang et al., 11 Feb 2026).

3. Asymmetric Advantage Policy Optimization

Asymmetric Advantage Policy Optimization (A2PO) is the mechanism used to combine semantic reward with business reward. Its central design choice is to treat the business reward as the primary anchor. In online deployment the anchor is eCPM; on public datasets, where monetary signals are unavailable, the paper uses a ranking proxy (Jiang et al., 11 Feb 2026).

The rejected baseline is naive reward summation, written as $$\pi_\theta(y\mid x)=\prod_{t=1}^{T}\pi_\theta(y_t\mid x,y_{<t}).$$3. The paper characterizes this strategy as brittle because the rewards have different scales, semantic reward may dominate, tuning $$\pi_\theta(y\mid x)=\prod_{t=1}^{T}\pi_\theta(y_t\mid x,y_{<t}).$$4 is sensitive, and updates can become unstable. A2PO therefore performs fusion in advantage space rather than reward space. For each sample it computes a business advantage $$\pi_\theta(y\mid x)=\prod_{t=1}^{T}\pi_\theta(y_t\mid x,y_{<t}).$$5 and a semantic advantage $$\pi_\theta(y\mid x)=\prod_{t=1}^{T}\pi_\theta(y_t\mid x,y_{<t}).$$6, both standardized within a rollout group. The fused advantage is

$$\pi_\theta(y\mid x)=\prod_{t=1}^{T}\pi_\theta(y_t\mid x,y_{<t}).$$7

with

$$\pi_\theta(y\mid x)=\prod_{t=1}^{T}\pi_\theta(y_t\mid x,y_{<t}).$$8

This gate is active only when business and semantic advantages have the same sign, and the paper notes the bound $$\pi_\theta(y\mid x)=\prod_{t=1}^{T}\pi_\theta(y_t\mid x,y_{<t}).$$9. Accordingly, semantic supervision can help but cannot overpower the business objective (Jiang et al., 11 Feb 2026).

The policy update uses a clipped policy-gradient objective,

$(x,a)$0

where $(x,a)$1 is the PPO-style importance ratio. An important empirical justification for this design is that directional consistency between business and semantic advantages is only about 45.4% on Amazon and 49.8% in the online ad system, which supports the need for consistency gating rather than unconditional fusion (Jiang et al., 11 Feb 2026).

4. Training Pipeline and Serving Architecture

The end-to-end workflow of S-GRec consists of five stages. First, PSJ is trained offline: the aspect-level scorer is trained with verified labels, and the aggregation model is trained with pairwise comparisons. Second, semantic rewards are generated sparsely, because PSJ is queried on only a small subset of training samples. Third, business and semantic advantages are computed from platform objectives and the PSJ holistic score, respectively. Fourth, A2PO gates the semantic advantage by consistency, fuses it in advantage space, and updates the generator with the clipped policy objective. Fifth, only the lightweight generator is served online, with no real-time LLM inference (Jiang et al., 11 Feb 2026).

The deployment section further specifies that the system uses business reward on 100% of data, PSJ on 5% of data, and an offline trainer periodically updates online serving parameters. This architectural decoupling is the core scalability mechanism of the framework (Jiang et al., 11 Feb 2026).

The implementation described in the paper uses Qwen3-4B for PSJ and a Qwen2.5-1.5B generator backbone with RQ-VAE tokenization. PSJ training data comprise approximately 20k point-wise aspect samples and 40k pairwise preference samples. The reported A2PO settings are learning rate $(x,a)$2, group size $(x,a)$3, and KL coefficient $(x,a)$4 (Jiang et al., 11 Feb 2026).

5. Empirical Evaluation and Industrial Deployment

Offline evaluation uses two Amazon review benchmarks, Industrial and Office, following MiniOneRec preprocessing: users and items with fewer than 5 interactions are filtered, there is no temporal or location metadata, and therefore the Context aspect is dropped offline. The baselines include GRU4Rec, Caser, SASRec, HSTU, TIGER, LC-Rec, BIGRec, D$(x,a)$5, S-DPO, and MiniOneRec. S-GRec is built on top of MiniOneRec, and the reported metrics are HR@3, HR@5, HR@10, NDCG@3, NDCG@5, and NDCG@10 (Jiang et al., 11 Feb 2026).

The paper reports the best performance on both datasets across all metrics. On Industrial, compared with MiniOneRec, HR@10 improves from 0.1586 to 0.1632, and NDCG@10 improves from 0.1167 to 0.1202. On Office, HR@10 improves from 0.1634 to 0.1689, and NDCG@10 improves from 0.1242 to 0.1308. The paper summarizes these gains as roughly 2.9% HR@10 and 3.0% NDCG@10 on Industrial, and 3.4% HR@10 and 5.3% NDCG@10 on Office (Jiang et al., 11 Feb 2026).

The ablations isolate both judge quality and fusion strategy. For PSJ, the aspect scorer improves from PairAUC 0.3987 / PointAcc 0.4043 with pretrained Qwen3-4B, to 0.4703 / 0.5202 with SFT, and to 0.8116 / 0.8687 with SFT+GRPO. For A2PO, the paper states that reward-level summation can significantly degrade HR/NDCG, whereas advantage-level fusion helps, and gating plus magnitude balancing improve stability further. Sparse semantic sampling is also emphasized: with only 5% semantic sampling, S-GRec reaches 99.1% of full performance on Office and 99.6% on Industrial, which the paper interprets as a 20× cost reduction (Jiang et al., 11 Feb 2026).

Industrial validation includes both simulation and online A/B testing. In simulation, S-GRec achieves 5.5% uplift in eCPM. In the online A/B test, deployed on 20% of WeChat Channels advertising traffic, it achieves +1.19% GMV, +1.55% GMV-Normal, +1.16% CTR, and -2.02% dislike rate. The abstract characterizes these as statistically significant gains in CTR and a 1.19% lift in GMV, obtained without requiring real-time LLM inference (Jiang et al., 11 Feb 2026).

6. Terminology, Scope, and Related Acronyms

A recurrent source of confusion is that the string “S-GRec” appears near several distinct recommendation and optimization methods. The exact titled method “S-GRec: Personalized Semantic-Aware Generative Recommendation with Asymmetric Advantage” is the framework centered on PSJ and A2PO (Jiang et al., 11 Feb 2026). By contrast, some summaries use “S-GRec” to refer to S$(x,a)$6GR, formally “Stepwise Semantic-Guided Reasoning in Latent Space for Generative Recommendation”, which is a different framework whose core mechanism is stepwise latent reasoning with thinking tokens inserted before each SID generation step (Guo et al., 26 Jan 2026).

The term should also be distinguished from S$(x,a)$7GRec, “Self-supervised Graph-based Point-of-interest Recommendation”, which addresses next POI recommendation in LBSNs through graph-enhanced self-attention and contrastive self-supervision (Li et al., 2022); from GRec, “Gap-filling based Recommender”, which uses partially complete sessions and a gap-filling encoder-decoder objective for session-based recommendation (Yuan et al., 2019); and from the unrelated optimization method SRG, “Stochastic Reweighted Gradient Descent”, which the provided summary notes is also discussed under an “S-GRec/SRG” naming pattern in stochastic optimization (Hanchi et al., 2021). It should likewise not be conflated with GREC, generalized referring expression comprehension in multimodal grounding, for which HieA2G is proposed (Wang et al., 2 Jan 2025).

Within generative recommendation specifically, the contrast between S-GRec and S$(x,a)$8GR is substantive rather than nominal. S-GRec uses an offline LLM-based semantic judge and business-anchored advantage fusion (Jiang et al., 11 Feb 2026), whereas S$(x,a)$9GR uses stepwise semantic-guided reasoning in latent space, CoBa RQ-VAE, and contrastively supervised thinking tokens aligned with coarse semantic clusters (Guo et al., 26 Jan 2026). The shared vocabulary of semantics, generation, and recommendation reflects topical proximity, but the architectural premises are different.