Unified Value Alignment for Generative Recommendation in Industrial Advertising DOI: XXXXXXX.XXXXXXX Conference: Make sure to enter the correct conference title from your rights confirmation email; June 03–05, 2018; Woodstock, NY ISBN: 978-1-4503-XXXX-X/2018/06 CCS: Information systems Recommender systems
Xinxun Zhang 1∗ , Yuling Xiong 2∗ , Jiale Zhou 3∗ , Zhengkai Guo 2 , Zhennan Pang 2 , Junbang Huo 2 , Jingwen Wang 2 , Xuyang Sun 2 , Enming Zhang 2 , Jiaguang Jin 2 , Changping Wang 2 , Yi Li 2 , Jun Zhang 2† , Xiao Yan 1 , Jiawei Jiang 1† , Jie Jiang 2† , · 2026
Read Unified Value Alignment for Generative Recommendation in Industrial Advertising DOI: XXXXXXX.XXXXXXX Conference: Make sure to enter the correct conference title from your rights confirmation email; June 03–05, 2018; Woodstock, NY ISBN: 978-1-4503-XXXX-X/2018/06 CCS: Information systems Recommender systems by Xinxun Zhang 1∗ , Yuling Xiong 2∗ , Jiale Zhou 3∗ , Zhengkai Guo 2 , Zhennan Pang 2 , Junbang Huo 2 , Jingwen Wang 2 , Xuyang Sun 2 , Enming Zhang 2 , Jiaguang Jin 2 , Changping Wang 2 , Yi Li 2 , Jun Zhang 2† , Xiao Yan 1 , Jiawei Jiang 1† , Jie Jiang 2† , (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-05803.txt.
Introduction
Advertising recommenders must please two masters: user interest and commercial value. Xinxun Zhang and fifteen co-authors present UniVA, a framework that aligns value through every stage of generative recommendation. Treating recommendation as next-token generation works for content feeds, but value signals usually get bolted on afterward.
Their tokenizer builds value-related attributes directly into item representations. A generation-as-ranking decoder then scores candidates with value in view, and the alignment continues into online serving. The system is presented as an industrial design, tested in the setting it was built for: large-scale advertising.
Catalog record
| CREATOR | Xinxun Zhang 1∗ , Yuling Xiong 2∗ , Jiale Zhou 3∗ , Zhengkai Guo 2 , Zhennan Pang 2 , Junbang Huo 2 , Jingwen Wang 2 , Xuyang Sun 2 , Enming Zhang 2 , Jiaguang Jin 2 , Changping Wang 2 , Yi Li 2 , Jun Zhang 2† , Xiao Yan 1 , Jiawei Jiang 1† , Jie Jiang 2† , |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 47 KB |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2605.05803
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Unified Value Alignment for Generative Recommendation in Industrial Advertising DOI: XXXXXXX.XXXXXXX Conference: Make sure to enter the correct conference title from your rights confirmation email; June 03–05, 2018; Woodstock, NY ISBN: 978-1-4503-XXXX-X/2018/06 CCS: Information systems Recommender systems Xinxun Zhang 1∗ , Yuling Xiong 2∗ , Jiale Zhou 3∗ , Zhengkai Guo 2 , Zhennan Pang 2 , Junbang Huo 2 , Jingwen Wang 2 , Xuyang Sun 2 , Enming Zhang 2 , Jiaguang Jin 2 , Changping Wang 2 , Yi Li 2 , Jun Zhang 2† , Xiao Yan 1 , Jiawei Jiang 1† , Jie Jiang 2† , Abstract . Generative Recommendation (GR) reformulates recommendation as a next-token generation problem and has shown promise in industrial applications. However, extending GR to industrial advertising is non-trivial because the system must optimize not only user interest but also commercial value. Existing GR pipelines remain largely semantics-centric, making it difficult to align value signals across tokenization, decoding, and online serving. To address this issue, we propose UniVA, a Uni fied V alue A lignment framework for advertising recommendation. We first introduce a Commercial SID tokenizer that injects value-related attributes into SID construction, yielding value-discriminative item representations. We then develop a Generation-as-Ranking SID Decoder jointly optimized by supervised learning and eCPM-aware reinforcement learning, which fuses value scores into next-item SID generation to perform generation and ranking in one decoding process. Finally, we design a value-guided personalized beam search that reuses generation-as-ranking logits as online value guidance and applies a personalized trie tree to constrain decoding to request-valid SID paths. Experiments on the Tencent WeChat Channels advertising platform show that UniVA achieves a 37.04% improvement in offline Hit Rate@100 over the baseline and a 1.5% GMV lift in online A/B tests. * * footnotetext: These authors contributed equally to this work $\dagger$ $\dagger$ footnotetext: Corresponding author 1. Introduction Driven by the generative modeling capability of large language models (LLMs) ( 1 ; 19 ; 31 ) , recommendation systems are shifting from multi-stage deep-learning pipelines to end-to-end generative architectures ( 13 ; 14 ) . Generative Recommendation (GR) formulates recommendation as a next-token prediction task by assigning each item a discrete Semantic ID (SID) and training an autoregressive model to generate SID sequences conditioned on a user’s interaction history ( 26 ; 29 ; 5 ; 23 ) . This formulation unifies user sequential modeling and item retrieval within a single framework, and has shown strong performance across search, e-commerce, and content recommendation ( 13 ; 21 ; 26 ) . Advertising recommendation is a core monetization task for internet platforms and is inherently multi-objective ( 12 ) . Unlike many other recommendation settings, it must jointly consider user satisfaction and monetization outcomes ( 16 ) . In practice, the system needs to balance user interest with advertiser-side signals such as bid, return on investment (ROI), and expected cost per mille (eCPM), so as to improve platform revenue without harming user experience ( 28 ; 27 ) . These requirements introduce a value-modeling problem that goes beyond semantic matching ( 26 ; 23 ; 22 ) . The success of GR across various industrial domains naturally raises the question of whether advertising can similarly benefit from this paradigm. Recent studies have shown encouraging progress in this direction. For example, GPR ( 26 ) introduces a unified tokenization scheme and model architecture for end-to-end generative advertising recommendation, while subsequent work further incorporates auction-aware objectives and business-side supervision ( 28 ; 23 ; 22 ) . However, existing approaches still optimize commercial objectives in a fragmented manner across different stages of the generative pipeline. Tokenization mainly preserves semantic similarity, autoregressive decoding remains dominated by generation likelihood, and online serving often depends on additional value ranking or filtering modules. Commercial value signals are therefore inconsistently modeled throughout SID construction, generation, and online serving, resulting in a mismatch between semantic generation and business objectives. We refer to this problem as value inconsistency in generative advertising systems, which mainly manifests in the following three aspects: ❶Value-insensitive SID tokenization. Existing RQ-based tokenization pipelines are primarily designed to preserve multimodal semantic similarity, but they do not explicitly model the commercial heterogeneity inherent in advertising ( 18 ; 15 ) . As a result, semantically similar advertisements may still correspond to substantially different monetization utilities, while being mapped into nearby SID paths. Such representations provide limited commercial discriminability within the token space itself, making downstream generation semantically coherent but insufficiently aligned with business objectives. ❷Semantic-dominated SID decoding. Existing GR methods mainly optimize commercial value at the training-objective level, while autoregressive SID decoding is still dominated by semantic likelihood and sequence consistency. This limitation is particularly critical in advertising recommendation because decoding directly determines which SID trajectories survive during token expansion. Once commercially promising paths are pruned at early decoding steps, subsequent objectives can no longer recover them. As a result, commercial value cannot directly influence autoregressive decoding decisions, causing commercially promising SID trajectories to be prematurely pruned during token expansion. ❸Value-unaware online serving. Even with value-aware training, online serving is still largely dominated by semantic beam expansion and heuristic filtering. Existing beam search procedures do not explicitly incorporate commercial value during SID trajectory selection, causing commercially promising paths to be overlooked during online decoding. Moreover, expanding over the full SID space wastes substantial computation on invalid candidates that violate inventory constraints or targeting rules, while introducing additional online value modules further increases serving latency and system complexity. Our Solution. The key insight of this work is that commercial value should not be treated as an auxiliary signal introduced after generation, but instead should be consistently embedded into SID construction, autoregressive decoding, and online serving within a unified generative framework. To this end, we propose UniVA , a Uni fied V alue A lignment framework for generative advertising recommendation, which propagates commercial value signals throughout tokenization, decoding, and online serving: • Commercial SID Tokenization. UniVA disentangles semantic organization and commercial refinement within hierarchical SID construction. Upper SID levels preserve semantic locality through RQ-based partitioning, while the final SID level specializes in commercial discrimination through a classify-then-bin strategy that integrates structured commercial attributes and bid discretization into value-aware token assignment. This design preserves semantic organization while improving commercial coherence among SID paths. • Generation-as-Ranking SID Decoder. UniVA introduces a dual-head Generation-as-Ranking SID Decoder that jointly models generative likelihood and commercial value. A generation head performs autoregressive SID prediction, while a value head estimates token-level commercial preferences within the same SID space. The two outputs are fused during decoding, enabling each autoregressive decision to jointly consider semantic relevance and monetization objectives. To achieve unified optimization, UniVA jointly trains the generation head with supervised learning and the value head with eCPM-aware reinforcement learning, allowing commercial value signals to directly shape SID generation trajectories during autoregressive decoding. • Value-Guided Personalized Beam Search. UniVA further proposes a value-guided personalized beam search for online serving. A request-specific valid-path trie tree first constrains beam expansion to feasible SID trajectories that satisfy inventory constraints and targeting rules. During decoding, beam scores are directly computed from the fused generation-ranking logits, allowing commercial value signals to participate in online SID trajectory selection. This design substantially reduces invalid search space while enabling