This work separates generation from selection: explanations are produced ahead of time as a frozen candidate pool (six prompt styles, two commodity LLMs), and a small CPU-resident selector picks one at request time, which needs no GPU and returns in under 100 ms.
Abstract
Industrial explainable-recommendation systems built on LLMs incur a substantial serving cost: each request triggers an LLM generation, with latency in the hundreds of milliseconds and cost that scales linearly with traffic. We separate generation from selection: explanations are produced ahead of time as a frozen candidate pool (six prompt styles, two commodity LLMs), and a small CPU-resident selector picks one at request time. The stack needs no GPU and returns in under 100 ms. Our primary benchmark is a 2,958-pair XRec Google Local subset, evaluating six offline-pool selectors (LambdaRank, PPO, GRPO, DPO, teacher-student distillation) and three KG-path selectors (random walks, edge-disjoint enumeration, MMR-reranked paths). A 300-pair MovieLens-1M split with Claude-Sonnet-4.5 references serves as an internal cross-dataset check, since no public benchmark exists for this setting. All variants use the same BERTScore-F1 protocol as XRec and G-Refer, averaged across five seeds. LambdaRank reaches F1 = 0.500 on Google Local, exceeding both G-Refer and XRec, and F1 = 0.329 on the MovieLens-1M check. With seed variance below 0.003 F1, the ordering is reliable: pairwise learning-to-rank outperforms single-action RL (PPO, GRPO, DPO), which use only one labelled candidate per rollout, leaving K-1 labels unused. The KG-path family targets a different objective: all three variants reach USR = 1.000 on Google Local and 0.997-1.000 on MovieLens-1M, since per-request path grounding yields a unique output per query, avoiding template-collapse failures affecting cached-LLM outputs. A generator-pool study comparing Claude 3 Haiku and Claude Haiku 4.5 shows small F1 shifts (0.001-0.006) while preserving selector ranking: selector and generator can be evaluated independently, though absolute F1 depends on the generator. End-to-end build cost is near $15 on commodity hardware.
Large Language Models (LLMs) have emerged as powerful tools for improving recommendation systems. The effectiveness of LLMs arises from their ability to harness rich textual information and their capacity to model heterogeneous user preferences based on users'interaction history. However, due to the large-scale and deep architectures, LLM-based sequential recommendation approaches generally incur high inference costs, resulting in a low return on investment. To mitigate this cost, many existing approaches resort to using only the first few tokens of item descriptions, which inadvertently discards valuable information contained in the full text, thereby leading to suboptimal recommendation performance. To address this limitation, we propose a novel Token Selection approach for Preference Optimization in LLM-based sequential Recommendation, i.e., TSPORec, which accurately pinpoints informative tokens throughout the entire textual content to improve recommendation performance. Specifically, we design a three-stage pipeline to select informative tokens and introduce a novel proxy reward to facilitate the implementation. TSPORec not only enhances recommendation performance but also improves computational efficiency. Extensive experiments across two models and datasets demonstrate the superb performance (up to 31.25%) and efficiency (up to 63.4%) of our approach compared with six baseline approaches. Code is available at https://github.com/WNQzhu/TSPORec.git.
Wenqiao Zhu, Chao Xu, Haipang Wu et al.· 0 citations
Modern recommender systems adopt Generator-Evaluator (G-E) for list-wise reranking: a generator produces sequences from candidates and an evaluator scores them at sequence-level to filter out the optimal one for exposure. Auto-Regressive(AR), working as the backbone for generative recommendation, suffers two limitations. First, its complexity grows linearly with list length, forcing the system to generate fewer lists under rigorous latency constraints and thus limiting exploration. Second, teacher-forcing creates a train-test mismatch; cumulative errors worsen with length and degrade quality. To address these problems, we propose Pair-Space Generation (PSG), a reformulation that elevates the generation atom from individual items to ordered item pairs. Given $n$ candidate items, PSG operates over pair vocabulary of size $n(n-1)$ per request, generates only $L/2$ tokens. Pair token representations are produced on-the-fly by a pretrained pair-token representation module optimized over large scale exposure logs, eliminating the data sparsity that would otherwise plague a quadratic sized vocabulary. We establish three theoretical guarantees: (i) PSG is bijective with item-space generation and induces an equivalent family of sequence distributions, thus incurring no loss of expressiveness; (ii) generation in pair-token space achieves approximately a $2\times$ to $4\times$ speedup theoretically under moderate settings and $1.83\times$ in the real industrial environmental settings; and (iii) under outcome-only rewards, the worst-case suboptimality of PSG is bounded by $O((L/2)^2 \bar{\epsilon})$, representing a nearly $4\times$ improvement over item-space generation. Beyond benchmark-based validation, PSG has also been deployed on Kuaishou, delivering a 0.178\% lift in per-user stay time on the platform, which serves over 400 million daily active users.
Chao Feng, Li Ma, Xiancheng Gao et al.· 0 citations
Online recommendation platforms increasingly use Large Language Models (LLMs) to extract structured features from ad creatives. While deploying a single-call LLM annotation agent yields significant Click-Through Rate (CTR) improvements in our live production environment, per-creative prompting is prohibitively expensive to scale. The redundant system instructions sent in every request account for 94% of billed input tokens. To break this cost bottleneck, we introduce PromptPack, a scalable, high-throughput LLM annotation agent. PromptPack achieves this scale via in-context batching, combining a shared system prompt, a strict XML structural envelope, and an output correction layer to ensure deterministic, pipeline-ready feature extraction across multiple creatives simultaneously. We evaluate PromptPack via an offline retrieval benchmark using a downstream logistic-regression ranker. To deeply profile the agent's behavior, we measure AUC and introduce Volume-Weighted Absolute Lift (VWAL), a novel metric capturing the signal quality of the generated features. Compared to our live, unbatched production baseline, PromptPack at batch size 20 cuts our LLM costs by 89% and accelerates throughput by 2.5x while fully preserving AUC.
Sebastian Koralewski, Merwan Barlier, Yulia Stolin et al.· 0 citations
Industrial recommender systems are commonly deployed as multi-stage cascades with separate candidate generators, pre-rankers, and final rankers. Although effective, these cascades require repeated user-history processing, complex feature pipelines, and multiple serving stages. Semantic-ID-based generative retrieval offers a path toward simpler end-to-end systems, but next-item prediction alone does not capture the fine-grained preferences encoded by production ranking objectives. We present Gryphon-v2, a unified generate-and-rank architecture for end-to-end recommendation. The model encodes a user history once, generates Semantic-ID candidates with an autoregressive decoder, resolves them to catalogue items, and ranks them with an item-level Ranking Module that reuses the shared encoder states. To transfer fine-grained production ranking preferences without adding an expensive second model to the serving path, we distill a high-capacity, training-only Teacher Ranker into the Ranking Module. Gryphon-v2 is trained with Rollout Distillation: teacher scores are the only ranking supervision, and they are collected over two complementary candidate distributions. Rollouts from the current decoder expose the Ranking Module to candidates produced by the same generation mechanism used at serving time, while logged impressions cover items users were actually shown. In an online A/B experiment on a large-scale recommendation surface at Yandex Music, a single Gryphon-v2 model replaces a production cascade comprising more than 15 candidate generators, pre-ranking, and final ranking. The deployment increases the number of active users by 1.41% at serving latency comparable to the production cascade. These results support the practical viability of a generative retriever with a Ranking Module distilled from the Teacher Ranker as an end-to-end alternative to a production cascade.
Anna Lipkina, Daria Tikhonovich, Viktor Yanush et al.· 0 citations
Modern industrial recommender systems have increasingly adopted the Generator-Evaluator (G-E) framework for the re-ranking stage. Within this paradigm, the generator produces candidate item lists from a pool filtered by upstream retrieval and ranking modules, while the evaluator scores these lists and selects the highest-scoring one for final exposure per request. However, on sequential platforms (e.g., short-video apps), users consume items continuously, ignoring artificial list boundaries. Conventional evaluators score lists by aggregating point-wise values, implicitly assuming exposure independence. This fails to capture critical session-level dynamics, such as contextual dependencies, user continuation, and diminishing marginal utility from repetitive content. To bridge this gap, we propose SWIM (Step-Wise Integrated Measure), a list-level evaluator that models user behaviors as a finite-horizon prefix session-level survival process. SWIM estimates the prefix-conditioned contribution of the current list to the session-level objective by factorizing it into a recursive survival distribution and reached-position conditional rewards. Leveraging a causally-masked Transformer, SWIM efficiently estimates continuation probabilities and utilities in parallel, satisfying strict industrial latency constraints. Extensive experiments demonstrate that SWIM significantly outperforms baselines in listwise reranking tasks, yielding substantial improvements in overall recommendation engagement.
Yuan Pu, Chenghao Zhang, Chao Feng et al.· 0 citations
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.
Xin Yu, Stephen Li, Sina Aghaei et al.· 0 citations
