S‑GRec: Personalized Semantic‑Aware Generative Recommendation with Asymmetric Advantage Alignment

Background and Motivation

In large‑scale online recommendation, autoregressive generative recommendation (GR) models show promise over traditional cascade architectures, but they rely solely on user behavior logs, leading to weak implicit intent signals and noisy supervision.

Although large language models (LLMs) provide rich semantic priors, industrial deployment faces two fundamental conflicts: high online inference cost and a mismatch between open‑domain semantic signals and platform business metrics such as CTR, eCPM, and GMV.

S‑GRec Overall Framework

To address these issues, the Tencent Advertising Technology team proposes S‑GRec (Personalized Semantic‑Aware Generative Recommendation), an end‑to‑end framework that decouples a lightweight online generator from an offline LLM‑based personalized semantic judge (PSJ). During training, carefully designed personalized semantic rewards and an Asymmetric Advantage Policy Optimization (A2PO) algorithm align deep semantic understanding with business value without increasing online latency.

PSJ: Two‑Stage Personalized Semantic Judge

PSJ transforms the complex user‑item semantic interaction into interpretable scalar rewards through a two‑stage pipeline.

Aspect‑level Semantic Scoring : For each candidate, the LLM evaluates four fine‑grained aspects—User Profile Relevance, Future Interest Relevance, Novelty, and Contextual Relevance—producing explicit scores.

User‑conditional Preference Aggregation : Because users weight these aspects differently, PSJ adopts a group‑related policy optimization (similar to GRPO) that learns pairwise preference weights from real historical feedback, yielding a stable personalized semantic reward.

A2PO: Asymmetric Advantage Policy Optimization

Simply adding semantic and commercial rewards causes severe multi‑objective conflicts. A2PO resolves this with three mechanisms:

Advantage Function Normalization : For each generated candidate set, commercial and semantic advantage functions are computed and group‑relative standardized to remove scale differences.

Asymmetric Consistency Gating : The commercial advantage serves as the anchor; semantic advantage is injected only when its direction aligns with the commercial direction, ensuring recommendations benefit both platform revenue and user experience.

Asymmetric Clipping Protection : When conflicts arise (e.g., high‑value items with low semantic advantage), A2PO automatically clips or suppresses the adverse semantic signal, protecting core business metrics.

Experimental Results

Evaluation was performed on Amazon open‑source benchmarks (Office and Industrial) and on a billion‑scale industrial recommendation system, using both offline metrics and online A/B tests.

Offline Benchmark

Compared with non‑LLM baselines (SASRec, HSTU, TIGER) and LLM‑based generative models (BIGRec, S‑DPO), S‑GRec achieved the best performance, especially improving novelty detection.

Ablation Study

Base + PSJ (traditional reward addition) caused a 44–63% drop in NDCG@10, confirming the target‑alignment conflict.

Base + PSJ + A2PO (full S‑GRec) yielded a 3–5% NDCG@10 gain, demonstrating A2PO’s ability to recover business performance while adding semantic value.

Semantic Guidance Analysis

When grouping test items by overlap with user history, S‑GRec’s HR@5 improvement grew with item novelty: +1.6% for re‑consumption items (Level 0) and +5.5% for unseen root items (Level 3), showing that PSJ’s world‑knowledge compensates for sparse behavior signals.

Online A/B Test

In production, S‑GRec increased GMV by 1.19%, CTR by 1.16%, and reduced dislike rate by 2.02% while preserving millisecond‑level latency. All LLM‑driven semantic computation occurs offline, so no GPU clusters are required online.

Conclusion and Future Work

S‑GRec demonstrates that generative recommendation can break the traditional cascade paradigm and that offline semantic judgment combined with asymmetric advantage alignment can inject deep LLM knowledge into a lightweight online engine without extra latency. Future directions include multimodal representation integration, long‑term intent tracing, and broader asymmetric alignment across scenarios.