Difficulty-Aware Semantic-ID Optimization (DASO), a tree-aware post-training method that addresses failure mode as an online rollout-allocation problem and improves over MiniOneRec-style GRPO on 11 of 12 metrics and achieves the best result on 9 of 12 metrics.
Abstract
Semantic-ID-based generative recommendation casts retrieval and ranking as autoregressive generation over hierarchical item identifiers. A common recipe is SFT followed by GRPO, yet vanilla GRPO is poorly matched to this tree-structured task. Under the frozen SFT checkpoint, the exact target is absent from the first 16 candidates of the 50-beam constrained ranking for many prompts, and in harder cases none of these candidates enters the target SID branch. This prompt-level diagnostic motivates a training concern: when on-policy GRPO groups are similarly target-missing, item-level rewards may produce weak or degenerate reward variation even if some candidates follow part of the target path. We propose Difficulty-Aware Semantic-ID Optimization (DASO), a tree-aware post-training method that addresses this failure mode as an online rollout-allocation problem. Instead of using fixed difficulty buckets or uniformly injecting ground-truth completions, DASO profiles each current rollout group by prefix-match depth, locates the bottleneck SID levels where candidates leave the target path, and reallocates a bounded portion of the group to prefix-guided completions while retaining raw rollouts for contrast. A SID-prefix reward provides graded credit, while an auxiliary SFT anchor mitigates regression on examples already solved by the SFT checkpoint. On the public benchmarks, DASO improves over MiniOneRec-style GRPO on 11 of 12 metrics and achieves the best result on 9 of 12 metrics; it also improves most level-wise recall metrics on the internal recommendation task.
Generative recommendation (GR) reformulates sequential recommendation as an autoregressive generation problem, where items are represented as discrete semantic IDs. However, learning effective item tokenization is critical yet remains challenging. Most existing methods optimize tokenization in a point-wise or heuristic manner, which often leads to Relational Structure Distortion. We identify two main sources of this distortion: (1) point-wise reconstruction losses that are insufficient to preserve item--item relations, and (2) collision-mitigation--driven code assignment that disrupts semantic neighborhoods under many-to-one mappings. This issue is especially harmful for recommendation, since the task fundamentally depends on relative item similarity. To address this issue, we propose ReSOT, a unified framework that Re-balances Semantic ID learning via Optimal Transport for GR. ReSOT consists of two complementary components: a structure-aware reconstruction module that aligns item–item relations between the original and the reconstructed embedding space, and an unbalanced optimal transport-based soft quantization module that achieves a better trade-off between collision mitigation and semantic neighborhood preservation. Extensive experiments on real-world datasets show that ReSOT provides a principled tokenization scheme that preserves relational structure while assigning codes in a collision-aware and semantics-consistent manner. Our code is available at https://github.com/grw-zju/ReSOT.
Renwu Geng, Yiming Xu, Fengxin Li et al.· Proceedings of the 32nd ACM...· 0 citations
Slate recommendation treats a slate rather than an individual item as the recommendation unit, requiring joint optimization of item interactions and slate utility. Existing approaches typically separate candidate generation from ranking and restrict optimization to retrieved candidates. Generative recommendation with Semantic IDs (SIDs) offers a path to end-to-end recommendation, but existing SID construction often lacks recommendation-aware semantics and effective local collaborative signals, while next-token prediction is misaligned with slate-level objectives. We propose OGR, an end-to-end framework that directly generates ordered slates-"Once Generated, Ranked."OGR first introduces TUSID, which adaptively fuses item-specific semantic and local collaborative information into hierarchical SIDs. It then uses list-wise preference planning and pipelined position-wise SID decoding to model global preferences and inter-item dependencies while generating ordered slates. We further propose SPA, a reward-guided conservative policy optimization method that aligns generated slates with user preferences beyond likelihood imitation. Offline experiments show that OGR outperforms representative baselines, with 48.2% and 27.2% relative NDCG@5 gains on industrial and public datasets, respectively. Online A/B testing on Kuaishou further yields a 1.120% improvement in Effective Views.
Generative recommenders select items by autoregressively decoding semantic identifiers (SIDs), whose token positions induce a coarse-to-fine hierarchy over the item space. In practice, SID decoders are trained via supervised next-token prediction, which imitates logged trajectories rather than directly optimizing downstream utility. This motivates post-training with outcome feedback to guide decoding toward higher utility. However, logged feedback is only observed for the final exposed item, causing most post-training methods to operate at the item level and broadcast the same terminal signal across all SID tokens. As a result, token-level credit assignment becomes sparse, high-variance, and layer-dependent. To this end, we propose Hierarchical Residual Policy Optimization (HRPO), a post-training framework that converts item-level outcomes into dense, token-aligned learning signals for conservative token-wise improvement. Specifically, HRPO first estimates SID prefix-level utilities via group-wise reward smoothing over feature-based user clusters. It then decomposes these utilities into residual token credits and accumulates them into credit-to-go signals. Finally, Residual-Return Policy Optimization (RRPO) optimizes the residual credits using clipped updates, group-normalized advantages, and KL regularization to preserve stability. Experiments on a public dataset and an online A/B test in a large-scale commercial system show consistent gains in session-level utility and key business metrics. Source code and the archived artifact are available for reproduction.
Kaifeng Guo, Yiming Yang, Jingtong Gao et al.· Proceedings of the 32nd ACM...· 0 citations
Semantic-ID generative recommenders represent each item as a short sequence of discrete semantic tokens and predict the next item by autoregressively generating this token sequence. This paradigm enables a unified generation interface for item IDs, histories, and item text, but it also creates a structured optimization bottleneck during reward-based post-training: when an early semantic token enters the wrong branch of the item-token space, finite rollout groups rarely reach the ground-truth item, so group-relative optimization receives identical zero rewards and produces no useful advantage. We propose Hint-Conditioned Generative Recommendation (HCGRec), a semantic-ID generative recommendation framework that recovers learning signal for such hard training instances. HCGRec diagnoses each instance with checkpoint rollouts and supplies a minimal target-prefix hint only when the current generator cannot reach the correct item. The model then generates the unhinted suffix under the hinted semantic branch, turning zero-reward groups into informative comparisons over item-token completions. Hinting also changes token identity: hinted prefix tokens are oracle-provided item context, while unhinted suffix tokens are sampled generation actions. We therefore introduce hint-aware credit decomposition, using supervised learning to preserve item-semantic and prefix-structure alignment for hinted tokens and GRPO to optimize the sampled suffix. Experiments on sequential recommendation benchmarks show that HCGRec substantially improves over supervised fine-tuning and vanilla reward-based post-training, while reducing zero-advantage training samples from over 70% to below 20%. The code is accessible at https://github.com/WncFht/GRec.
Kangning Zhang, Hao Fang, Xukun Luo et al.· 0 citations
Gwhere is proposed, an end-to-end industrial framework that integrates semantic identifier (SID) generation with LLM-based generative next POI recommendation and adapts LLMs to mobility scenarios via continued pretraining on enriched spatio-temporal corpora, supervised fine-tuning, and Exposure-Aware Kahneman-Tversky Optimization.
Penglong Zhai, Bowen Zheng, Jie Li et al.· 0 citations
Generative Recommendation (GR) has emerged as a new paradigm for sequential recommendation, in which a representative line of work encodes items into hierarchical semantic IDs via residual quantization and predicts the IDs token by token. However, this generative formulation still exhibits structural gaps with respect to the recommendation task: flattening multi-token IDs into a single sequence destroys item-level structure, and the inconsistency between training and inference over a hierarchical codebook gives rise to semantic drift. To bridge these two gaps, we propose BARGE, which employs Item Context-Aware Attention (ICA) to restore item-level structure during encoding, and Hierarchical Path Reranking (HPR) together with Dual-Path Decoding (DPD) to suppress semantic drift from two complementary angles during decoding. Extensive experiments and analytical studies on public benchmarks and a large-scale offline test demonstrate that BARGE achieves superior recommendation performance. An online A/B test on a Tencent platform yields improvements of 0.60% in click-through rate, 1.34% in click unique visitors, and 1.70% in total reading time, confirming the practical value of BARGE in industrial-scale recommendation.