Once Generated, Ranked: End-to-End Generative Slate Recommendation with Unified Semantic-Collaborative IDs

Abstract: 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.

• The paper introduces OGR, combining TUSID identifiers, GL2P list-wise planning, and SPA preference optimization to achieve up to 48.2% higher NDCG@5 than strong baselines and a 2.49× throughput gain.
• TUSID integrates multimodal semantics with co-occurrence signals through confidence-aware weighting, improving Effective Recall@5 to 0.148 while maintaining collision-free identifiers and strong semantic cohesion.
• GL2P reduces decoding depth from O(KD) to O(K+D), while conservative SPA alignment improves slate-level preferences and supports online gains in effective views, comments, likes, and forwards.

Motivation and problem setting

Slate recommendation treats an ordered list of items, rather than a single item, as the unit of prediction, requiring joint modeling of inter-item interactions and list-wise utility. The paper identifies two structural deficiencies in existing systems. First, cascaded generate–rank–rerank pipelines bound slate quality by the retrieved candidate pool and prevent end-to-end optimization because objectives are misaligned across stages. Second, although Semantic ID (SID)-based generative recommendation offers a path to end-to-end generation, prevailing approaches optimize next-token prediction (NTP) over individual SID tokens, which is misaligned with slate-level utility and does not explicitly capture cross-position dependencies. The proposed framework, OGR ("Once Generated, Ranked"), addresses both by directly generating ordered slates conditioned on user history, unifying item-level generation with list-wise organization.

TUSID: unified semantic–collaborative identifier construction

The first contribution is Two-stage Unified SID Construction (TUSID). In the semantic stage, a multimodal LLM encodes image, audio, and text into a fine-grained representation $e_i^{\mathrm{M}}$, which serves as the query in cross-attention over four high-level attribute embeddings (Brand, Category, supply-side B-tags, demand-side C-tags). A sigmoid gate regulates the attribute residual added to the MLLM backbone, and fusion parameters are trained with a SASRec-style next-item loss plus an $\ell_2$ regularizer on the residual to prevent the attribute branch from overwhelming the backbone. This asymmetric design is motivated by the observation that naive additive fusion of heterogeneous semantic sources degrades downstream performance relative to gated cross-attention.

In the collaborative stage, rather than aligning representations via contrastive sampling, TUSID injects local co-occurrence structure before quantization. CountSketch-Based Collaborative Embedding (CCE) compresses each item's distance-weighted contextual distribution ($1/|p-q|$ within window $\omega$) into a fixed-dimensional sketch using shared bucket and sign hashes; because hashes are shared across items, sketch inner products unbiasedly estimate raw co-occurrence inner products. Signed-log transformation, normalization, and random projection yield $e_i^{\mathrm{col}}$. Confidence-Aware Weighting (CAW) then scales collaborative contribution by distinct-user support $U_i$ through $\gamma_i = \log(1+U_i)/(\log(1+U_i)+\tau)$, so cold-start or sparse items fall back entirely on semantics—consistent with industrial practice. The unified representation is a norm-preserving confidence-weighted concatenation quantized by RQ-KMeans into four-level SIDs with 1,024 codes per level.

Against six public SID methods under the same OGR backbone, TUSID achieves the best downstream metrics (Effective Recall@5 of 0.148 vs. 0.129 for the strongest baseline) while remaining collision-free (ICR/CUR of 1.000) and preserving semantic cohesion (V-measure 0.424, SC 0.747). Notably, LETTER attains strong codebook balance but the worst codebook utilization (CUR 0.335), supporting the paper's claim that effective SIDs require balancing semantic preservation, collaborative structure, and codebook quality rather than optimizing any single criterion.

GL2P: list-wise planning with pipelined decoding

The Global Listwise Preference Planning architecture (GL2P) replaces flat NTP over concatenated slates with a two-module decomposition. A causally masked preference planner autoregressively produces one planned preference representation per slate position from encoded user history, capturing cross-position dependencies in a coarse-grained space. A position-wise SID decoder then generates each position's four-token SID chain conditioned on its planned embedding, preceding intra-item tokens, and the history. Because cross-position dependencies are delegated to the planner, SID decoding at position $m$ can overlap with planning for position $m+1$, reducing sequential dependency depth from $\mathcal{O}(KD)$ to $\ell_2$0—an idealized critical-path reduction factor of about 2.22 for $\ell_2$1, $\ell_2$2. Measured throughput gains are $\ell_2$3 and $\ell_2$4 over beam-search TIGER and OneRec respectively, while achieving the highest NDCG@5; Top-5 selection variants are faster still but substantially less accurate because they ignore ordering dependencies.

Supervision combines two branches: an exposure-order sequence preserving logged display distributions and a feedback-reordered sequence injecting preference signals, weighted by $\ell_2$5. Ablations show removing either branch hurts performance, and removing the list-wise planner halves Effective Recall@5 (0.0416 → 0.0201) even when its parameters are reallocated—strong evidence that global planning, not capacity, drives the gain.

SPA: slate-level preference alignment

Supervised generation remains likelihood imitation of fixed targets, so Slate-Level Preference Alignment (SPA) performs post-training with PPO-style clipped policy optimization over candidate slate actions. Rewards comprise a primary component aggregating feedback signals (effective view +0.10, completion +0.15, like +0.20, share +0.15, skip −0.15, dislike −0.25) and an auxiliary component dominated by intra-slate diversity (0.90) with novelty (0.10). Reward-Consistent Calibration standardizes both rewards within each user's action set and gates the auxiliary term by an agreement coefficient that suppresses it when primary and auxiliary scores conflict in sign. Conservative regularization—a KL penalty between reference and current candidate-slate distributions plus supervised replay—constrains policy drift; only the planner and SID decoder are updated. Removing this regularization causes the largest single ablation drop (NDCG@5 0.0633 → 0.0312), indicating that unconstrained alignment destabilizes generation. Removing SPA alone costs roughly a third of NDCG@5, confirming that likelihood training alone cannot optimize relative preferences among generated slates.

Experimental results

On KuaiRec and a proprietary Kuaishou dataset with leave-five-out evaluation, OGR outperforms sequential recommenders (Caser, SASRec, BERT4Rec), listwise rerankers (PRM, Seq2Slate), and generative recommenders (TIGER, OneRec), all at comparable parameter budgets (~32 MB). Relative NDCG@5 improvements reach +48.2% on the industrial dataset (0.0633 vs. 0.0427 for Seq2Slate) and +27.2% on KuaiRec (0.1303 vs. 0.1024 for PRM). An instructive negative result is that OneRec achieves the best Hit Rate/Recall among baselines but lower NDCG than listwise rerankers, exposing what the authors call a decoding dilemma: autoregressive slate generation captures dependencies but suffers serial latency and error accumulation, whereas probability-based top-$\ell_2$6 selection is efficient but not listwise-aligned. Online A/B testing on 3% of Kuaishou production traffic over one week yields gains of 1.120% in Effective Views, 2.954% in Comments, 0.505% in Likes, and 1.255% in Forwards.

Limitations and open questions

Several constraints qualify these results. First, the offline protocol trains on logged exposure slates, so exposure-order supervision inherits presentation bias; the feedback branch mitigates but does not eliminate this, and counterfactual slate rewards are unavailable—SPA's action sets consist only of logged exposure slates evaluated under the policy, not newly generated ones, which limits the exploration scope of alignment. Second, the reward schema uses fixed hand-tuned weights and, due to confidentiality, cannot be fully disclosed, leaving sensitivity to weight choices unexamined. Third, CCE relies on positive-feedback prefixes and a fixed context window ($\ell_2$7); how distance-weighted sketches behave under different interaction densities or longer horizons is not analyzed. Fourth, the efficiency analysis reports throughput against beam-search baselines at a specific configuration ($\ell_2$8, $\ell_2$9); scaling behavior at larger slate sizes or deeper SID hierarchies remains open. Finally, whether the listwise planner's coarse preference space generalizes beyond short-video consumption signals is untested.

Conclusion

OGR demonstrates that generative slate recommendation can be made both accurate and efficient by pairing recommendation-aware SID construction (TUSID) with decomposed list-wise planning and pipelined decoding (GL2P), followed by conservatively regularized slate-level preference alignment (SPA). The combination yields consistent offline gains across industrial and public benchmarks, a structural reduction in decoding depth from $1/|p-q|$0 to $1/|p-q|$1, and verified online impact at production scale. The principal open questions concern off-policy reward calibration, reward-weight sensitivity, and scalability of the planning–decoding pipeline beyond the evaluated configurations.