e-commerce platform engineers

Abstract: Recommenders on Web platforms such as e-commerce and music streaming need their training recipes re-tuned for each dataset. Large language model (LLM) agents can now rewrite these recipes, enabling recursive self-improvement (RSI) of recommender systems. Existing agents propose one edit per round, fully retrain the recommender, and keep the edit only if its ranking quality improves. Unlike programs or patches that can be scored by execution, a recommender edit changes how the model is trained, which raises two challenges. First, ranking feedback is delayed: an edit's ranking quality is known only after full retraining, so much of the retraining budget goes to edits that are later discarded. Second, the signals available before retraining are unreliable: sampled training losses can look solved while full-ranking performance remains poor, and they are not comparable across edits that change the loss or sampler. We propose RecRSI, which generates diverse candidate edits from a diagnosis of the current model and selects one for full retraining using early ranking feedback: each candidate receives a short trial, the first steps of its full training run, scored by full-ranking NDCG@10 on held-out probe users. With two sequential recommenders on three datasets and the same number of full retrainings, RecRSI significantly outperforms existing self-evolving agents in every setting, and on both Amazon datasets reaches in its first round a ranking quality that none of them reaches in five.

Abstract: Long user behavioral sequences improve the performance of generative recommendation (GR), but computing full attention over long histories during autoregressive beam search decoding incurs substantial online inference latency. To accelerate long-sequence decoding, existing approaches typically rely on static history compression or dynamic top-k sparse attention; however, the former risks losing critical semantics, whereas the latter incurs significant overhead from repeated token selection and non-contiguous memory access. In this work, we thoroughly analyze attention patterns during GR decoding and reveal two key properties: (i) progressive attention focusing, where attention progressively focuses on a small subset of critical historical tokens; and (ii) cross-step ranking consistency, where their importance rankings remain relatively stable across subsequent decoding steps. Motivated by these observations, we propose Progressively Focused Attention (PFA), an efficient sparse decoding framework for long-sequence GR. PFA retains full attention at the initial decoding step to explore global historical preferences. It then establishes a shared importance ranking through cross-beam weighted aggregation and reorders the key-value cache of the selected tokens into a contiguous memory layout. Subsequent decoding steps reuse this ranking and geometrically reduce the token budget, preserving critical semantics while lowering computational and memory access overhead. Extensive experiments on two public benchmarks and a large-scale industrial dataset demonstrate that PFA outperforms state-of-the-art baselines. A 7-day online A/B test on a large-scale e-commerce advertising platform further confirms that PFA delivers substantial business gains while maintaining low inference latency. Our code is publicly available at https://anonymous.4open.science/r/PFA-review-088D.

Abstract: Coordinating multiple recommendation objectives (e.g., click and conversion) is critical for real-world generative recommendation (GR). However, since a positive item in GR can be retrieved only when its tokens are successively selected at every decoding step, traditional item-wise methods lack fine-grained token-wise supervision to guide intermediate steps. Motivated by its success in foundation models, we consider introducing multi-objective on-policy distillation (MOPD) into GR to distill two single-objective teachers into a student at each token. Yet, we find that in vanilla MOPD, (i) teachers exhibit unequal distillation difficulty at each token (i.e., token-level imbalance), and (ii) the harder-to-distill teacher varies across tokens (i.e., cross-token inconsistency), leaving each objective poorly learned at certain steps and hindering positive item retrieval for both objectives, leading to suboptimal performance. To address this challenge, we propose KL-MOPD (KL-Weighted Multi-Objective On-Policy Distillation), a simple yet effective distillation method that allocates token-wise teacher weights proportional to the initial student-teacher token-wise KL divergences, thereby strengthening supervision for the harder-to-distill teacher at every decoding step. Despite its simplicity, we theoretically prove that KL-MOPD strictly outperforms vanilla MOPD in the additive joint recommendation objective under mild conditions. Moreover, KL-MOPD is plug-and-play and computationally efficient, requiring only a one-time pre-training weight calibration with negligible training overhead. Extensive experiments on three public benchmarks, a billion-scale industrial dataset, and online A/B tests on our e-commerce advertising platform demonstrate that KL-MOPD significantly improves both individual and joint recommendation performance. The anonymous code is provided for peer review.

Abstract: Generative recommendation (GR) represents items as semantic identifiers (SIDs) and autoregressively generates the next item's SID from interaction history. However, most methods assign the same SID length to all items despite highly unequal interaction frequencies. This uniform design repeatedly incurs the cost of fine-grained codes for popular items, while shortening codes to reduce this cost also compresses the semantic representations of long-tail items. Inspired by the compact forms of frequent expressions in natural language, we propose JevRec to encode items as variable-length words and generate the next item from interaction context. Linguistic encoding retains frequency-dependent semantic prefixes from shared codebooks, producing shorter SIDs for popular items and richer descriptions for long-tail items. Yet predicting a sequence-level length and generating token-level content impose different optimization requirements that may conflict under a shared output head. Our Jev-like decision modeling component therefore predicts a length distribution with a dedicated head and generates content conditioned on length through a separate head while sharing the interaction-history representation. Experiments on three public datasets show that JevRec outperforms all compared baselines across the evaluated recommendation metrics. Further analyses show lower training and inference time than TIGER and reduced gradient conflict between length and content prediction.

Abstract: Sequential recommendation models typically predict a user's next action from recent session interactions, which limits personalization when short-term signals are weak or ambiguous. Rather than assuming that incorporating longer-term history and community signals is universally beneficial, we study when it actually helps. We propose BehaveRec, a three-signal framework combining a causal session transformer, a bidirectional User History Encoder over a 90-day window, and a GraphSAGE module over a user-user graph built from behavioral embedding similarity rather than item co-occurrence; a learned gate blends the three signals per user. On an industrial dataset of 137M sequences from 12.6M users, BehaveRec improves R@5 by +3.2% over SASRec, with a larger-capacity SASRec baseline (more parameters than BehaveRec) still trailing by +2.5%, ruling out capacity as the explanation. More importantly, a user-level stratification by history length reveals a clear dose-response relationship: BehaveRec's advantage is negligible below roughly 10 interactions and grows monotonically with history richness, reaching +0.017 R@5 for users with 101+ events. This pattern is corroborated across one industrial and three public datasets spanning history-rich to history-sparse regimes, where session-only models remain preferable. These headline results, and the resulting deployment threshold, are established for category-level intent prediction; a preliminary, single-run check on the full item-level vocabulary yields a directionally mixed result and is scoped as an open question rather than a claim. We use these results to derive a practical deployment rule -- roughly 20 interactions per user -- for when long-term history and behavioral-graph signals are worth their added complexity. Despite this added complexity, BehaveRec remains lightweight: it uses 5.67M parameters and trains on a single GPU in under two days.

Abstract: Cross-domain sequential recommendation aims to improve next-item prediction in a target domain by leveraging user behavior from auxiliary source domains. However, real-world deployment faces two critical challenges: business constraints that isolate interaction data across domains, and user heterogeneity, where many users lack cross-domain activity. This makes existing cross-domain methods difficult to apply. In this paper, we propose a simple and practical framework for cross-domain sequential recommendations that transfers information exclusively through learned user representations, without sharing raw interaction data or item embeddings across domains. We introduce a lightweight adapter that projects the source-domain user representation into the target-domain embedding space and integrates it with the target interaction sequence. Importantly, when source data is unavailable (i.e., the user observed only in the target domain), the model seamlessly defaults to a robust target-only sequential recommender, retaining benefits from joint training across all users. Extensive experiments were conducted under two setups; one targets benchmarking against previous baselines, while the other focuses on realistic and scalable conditions. Both sets of experiments demonstrate consistent performance gains for shared and unshared users.

Abstract: User memory helps personalized systems build on their understanding of users over time. In e-commerce, users convey their preferences and needs through both explicit language and implicit interactions with products. We study memory built from logs of these interactions, examining how it can \emph{remember} concrete evidence and established interests, and use that evidence to \emph{infer} underlying needs and preferences that inform judgments beyond well-observed product domains. Existing memory methods primarily target language-based interactions, and evaluations offer limited real behavioral data and limited separation of these capabilities. This leaves open how to jointly support and separately evaluate remembering and inference from behavioral logs. We formulate the problem of continual behavioral user memory around these joint requirements and develop a dedicated evaluation framework and memory method to address them. \textbf{BehaviorMemBench} is a benchmark built from real e-commerce logs covering 1{,}194 users and 17{,}462 behavioral batches over more than three years, with automatic future-behavior labels and a task-independent memory interface to separately evaluate remembering and inference under a common protocol. \textbf{BehaviorMem} is a memory design that organizes explicit behavioral facts and inferred user characteristics across long-term and short-term scopes in a two-axis state, with evidence-grounded formation, continual maintenance, and query-independent readout. Experiments across two builders and two task readers show that BehaviorMem leads the evaluated agentic memories on remember and infer, and reveal different method rankings across the two capabilities and competitive statistical baselines. Validation on a public dataset further supports its usefulness for downstream tasks based on behavioral histories.

Abstract: Generative bundle construction completes a partially observed bundle by generating its missing items from a large product catalog. Compared to autoregressive LLMs, diffusion LLM emerges as a natural fit for this unordered generation task. Their bidirectional attention provides an inductive bias better aligned with unordered bundle construction. Iterative denoising can reveal multiple tokens per step, potentially reducing inference latency. However, directly applying DLMs to this task presents two key challenges: 1 Despite its flexible reveal order, confidence-guided decoding tends to postpone semantic-fork decisions until refinement tokens constrain their choices, reducing bundle diversity. 2 Revealing more tokens per step can compromise recommendation quality, stifling the potential for faster inference. We introduce DPE-Rec, a diffusion framework for diversity preservation and efficient recommendation. DPE-Rec preserves diversity by introducing semantic-fork-aware GRPO, which prioritizes semantic-fork decisions via autoregressive rollouts. It improves efficiency through pseudo-trajectory distillation, which concentrates learning from the vast random corruption space onto the compact trajectory-induced subspace visited by high-quality decoding, unlocking reliable multi-token decoding. Together, these components enable diverse and accurate bundle construction with substantially fewer denoise steps. Experiments on multiple bundle construction benchmarks show that DPE-Rec improves recommendation performance by 63% on average, and its fast variant achieves 25.6$times$ faster inference on average than existing generative recommendation arts, with minimal loss in recommendation quality.

Abstract: Entity matching is the task of deciding whether two records describe the same real-world entity. Current entity matching methods rely on pre-trained language models (PLMs) which require task specific fine-tuning or generative large language models (LLMs) which do not require task-specific fine-tuning but are slow and expensive at matching large record sets. Decision models are a new class of LLM-based models that return a probability for each predefined answer option rather than generate text. This paper investigates decision models as an alternative to generative LLMs and fine-tuned PLMs for entity matching. We evaluate six decision models in a zero-shot setting across eight benchmarks and compare their performance with that of the generative LLM GPT-6 Luna and the PLM-based matcher Ditto. Our experiments show that, without task-specific training data, the four strongest decision models achieve mean F1 scores close to that of GPT-6 Luna, with the best model slightly surpassing it. This model also outperforms Ditto on six of the eight benchmarks. All decision models are faster than GPT-6 Luna, and most are less expensive. For three of the four strongest decision models, the predicted match probabilities are well calibrated without requiring any task-specific recalibration. A model cascade that routes the least confident record pairs to an ensemble of generative LLMs improves the mean F1 scores of the four strongest decision models by 1.9-4.0 percentage points. Using the best decision model and its best-performing prompt for each benchmark, the cascade surpasses the highest published F1 scores from literature on four of the eight benchmarks.

Abstract: Large language models (LLMs) bring world knowledge to user preference modeling in sequential recommendation, but generation-based recommenders express next-item predictions indirectly as text. Generated outputs may fail to identify a unique candidate item, depend on candidate presentation order in the prompt, and lack explicit confidence in the recommendation. Recent studies explore Jev, a model for fast option selection with explicit confidence. Inspired by Jev, we introduce decision-native recommendation, a new paradigm that jointly models item choice and recommendation confidence using LLM representations. It requires valid candidate selection, consistency under candidate reordering, and confidence that reflects recommendation accuracy. We propose Direct Item Choice Estimation (DICE), which learns a personalized next-item distribution over supplied candidates from user interaction history and recommends the highest-scoring item with calibrated confidence. Tree-structured attention shares history computation across isolated candidate branches for order-consistent candidate scoring in a single forward pass. On three recommendation datasets, DICE achieves higher top-one accuracy than all evaluated baselines. Separate evaluations show unchanged item choices under tested candidate reorderings and end-to-end speedups of up to approximately $6\times$ over the evaluated generation-based methods. Its confidence estimates support selective recommendation by identifying more reliable predictions and guide additional computation for low-confidence examples in offline routing.

Abstract: Geo-filtered approximate nearest-neighbor (GFANN) search retrieves the objects inside a query-specified geographic disk whose feature vectors are closest to a query vector, a core operation in location-aware Web search. A disk filter on a standard proximity graph invalidates out-of-disk routing vertices; on selective disks, these removals break greedy paths and recall drops sharply. Existing methods scan all in-disk objects, over-fetch unfiltered candidates, or maintain separate sub-indexes across spatial partitions, trading latency for index overhead. We propose REACH, a single-graph index based on row-local active-order invariance: if every vertex stores its outgoing neighbors in a fixed-width row sorted by vector distance, any disk filter masks out-of-disk entries while surviving in-disk (active) neighbors remain vector-sorted. Offline, each vertex merges globally closest vector candidates with a nearest-first geographic stream into one row; online, beam search evaluates vector distances only for the first l active entries of each visited row. Because co-located objects share the same distance to every query center, every disk admits a coordinate group in full or not at all, decomposing active connectivity into intra-group and inter-coordinate conditions. Under a spatial marked-site model with repeated coordinates, we prove that polylogarithmic out-degree guarantees strong connectivity across all disks and constant-factor distance reduction per hop on compact disks, while a cut-rank threshold characterizes the minimum scan budget l. Across four datasets up to 10^7 objects, one REACH index per dataset outperforms SOTA methods, achieving 2.49x higher geometric-mean QPS than the strongest baseline at 0.95 Recall@10 on selective queries, while building 3.8x-11.9x faster and using 37%-45% less space than a SOTA spatial-graph index.

Abstract: Emerging e-commerce platforms, such as live-streaming and Online-to-Offline (O2O) services, inherently operate under a "Container-based" display paradigm. In this setting, a critical granularity mismatch exists: the platform performs ranking and estimation at the container level (e.g., livestream rooms or stores), whereas the actual transactions occur at the item level within these containers. This structural discrepancy complicates Gross Transaction Value (GTV) estimation, introducing unique challenges including high label variance, severe label sparsity, and dynamic item heterogeneity. To address these challenges, we propose a novel framework named Hierarchical GTV Estimation (HGE). HGE employs a Hybrid Set-Aware Encoder (HSAE) to model the dynamic composition of items and feature interactions. To mitigate label variance, we introduce a Proxy-Label Learning (PLL) strategy that decomposes post-click GTV into purchase probability, quantity, and unit price. Furthermore, a Residual Correction and Ensemble Fusion (RCEF) module is designed to handle label sparsity by effectively combining item-level aggregation with container-level inference. Extensive experiments on large-scale industrial datasets demonstrate that HGE significantly outperforms state-of-the-art baselines, achieving a +0.63% lift in XAUC offline and a +1.42% increase in Revenue Per Search (RPS) in online A/B tests. HGE framework has been fully deployed into our main traffic.

Abstract: Generative recommenders using semantic identifiers predict items through successive token decisions, but beam search may irreversibly prune promising item paths. This limitation is particularly consequential when a few recent interactions signal an emerging interest while earlier behavior favors established preferences. Yet an independently trained discriminative recommender may rank the same items highly from the same interaction history. Such complementary scores are available at test time, offering a way to adapt a frozen generator without retraining either model. The challenge is to use these scores before pruning: combining completed recommendation lists cannot alter earlier beam decisions, while scores over complete items cannot directly inform choices among semantic prefixes shared by multiple items. We propose \textbf{T}est-Time \textbf{C}ollaborative \textbf{B}eam \textbf{S}earch \textbf{A}daptation (\textbf{TCBSA}), a training-free method that incorporates collaborative information into semantic decoding. TCBSA transforms discriminative score margins into potentials over shared semantic edges and uses a confidence gate to regulate their contributions to cumulative path scores before beam selection. We establish a sufficient condition for target prefix survival under global beam competition and bound score perturbations under low discriminative confidence. Full-catalog experiments show consistent improvements over strong baselines across datasets and different generative and discriminative recommenders. Analyses of emerging interests and decoding behavior demonstrate gains in recommendation accuracy and target path survival, while targeted interventions support the contribution of shared prefix corrections. TCBSA also offers a favorable balance between recommendation accuracy and inference latency compared with simply widening the native beam.

Abstract: Generative recommenders replace the retrieve-then-rank cascade with a single model: beam search over semantic item identifiers performs retrieval, and the same likelihood performs ranking. We show that this shared score makes them forget long-tail items twice. In retrieval, beam search keeps identifier prefixes by the probability mass of their subtrees, so tail items are pruned early, and a wider beam admits few more of them. In ranking, the likelihood places tail items below head items even when they are admitted, so expanding retrieval alone brings little or no gain. Both failures reflect missing item-level memorization (i.e., direct recall of which items co-occur), which collaborative filtering provides and shared-prefix generation lacks. To this end, we propose CAPE (Co-occurrence-Aware Pool Expansion), a training-free plug-in that restores memorization at both stages: (i) a memorization channel retrieves co-occurrence-supported tail items outside the beam and scores them with the frozen generator, and (ii) memorization-calibrated ranking fuses the generator likelihood with co-occurrence evidence while keeping the generator's first item. Experiments with two generative backbones on product and music domains show that CAPE significantly improves long-tail accuracy over the backbones and over re-ranking baselines that share its candidates, preserves or improves overall accuracy, and adds a small fraction of the latency.

Abstract: The growing use of autonomous artificial intelligence agents in digital commerce creates new security concerns involving agent impersonation, unauthorized payments, compromised agents, and misuse of delegated authority. This research introduces the Autonomous Commerce Agent Guard System (ACAGS), a cybersecurity framework combining supervised machine learning, behavioral analytics, agent-specific authorization controls, and explainable AI. The framework was evaluated using synthetic financial transaction data augmented with simulated agent characteristics and five adversarial scenarios. XGBoost was selected following comparisons with alternative classifiers. Across ten experimental runs, behavioral features improved mean F1-score from 52.12% to 67.49%, with agent-specific features increasing performance to 69.46%. The framework achieved detection rates of 99.77% for abnormal agent behavior, 99.73% for impersonation, and 100% for malicious delegation and authorization-scope violations. Compromised-agent detection reached 72.27%. ACAGS also supported APPROVE, REVIEW, and REJECT authorization outcomes, with SHAP explanations providing model interpretability. The results demonstrate the potential of combining behavioral intelligence and delegated authorization controls to strengthen simulated agentic-commerce payment security, while highlighting the need for validation using real-world transactions.

Abstract: Generative recommendation predicts items by autoregressively generating semantic identifiers (SIDs), whose tokens are shared across semantically related items. With this paradigm gaining adoption, the question of how to support unlearning in it has become pressing. If training interactions must later be removed, such as injected spam or interactions with items a user no longer wants, this sharing lets an unlearning update unintentionally affect the predictions for the interactions' neighbors. We show that this effect is locally structured and hidden by global metrics: unlearning can leave recommendation utility on the whole test set nearly unchanged while substantially degrading it on neighborhoods of the forgotten interactions' SIDs. We propose Neighborhood-Aware Unlearning (NAU), which augments interaction-level unlearning with explicit control over the probability mass assigned to the forgotten items' neighborhoods. Across three datasets and four SID tokenizers, NAU restores spam-induced exposure toward the retrained reference while retaining more than 95% of retrained utility, and recovers the neighborhood utility that forget-and-repair unlearning loses at no global cost. We identify local probability redistribution as an important failure mode of unlearning in SID-based generative recommenders and show that explicitly controlling this redistribution can mitigate its effects.

Abstract: Generative retrieval assigns each item a discrete semantic identifier (SID) and retrieves items by autoregressively generating SID sequences. In dynamic environments such as e-commerce, new items arrive continuously, requiring periodic updates to the SID-to-Item mapping and model to maintain retrieval timeliness. Each update, however, can invalidate search paths already learned by the retriever, and the quantitative relationship between mapping change and retrieval quality has not been systematically characterized. We study this problem with extensive controlled experiments over 67 days of large-scale e-commerce search logs and obtain three core findings. (1) SID semantic drift causes severe retrieval degradation that may be masked by large beam sizes: mapping updates reduce HitRate by 41.78-62.05% at beam size 1 but only 8.33-10.85% at beam size 1000, as redundant candidate generation conceals the true severity of drift. (2) The fraction of changed SIDs is strictly monotonically related to retrieval quality: as the remapped fraction increases from 0% to 72.75%, HitRate decreases monotonically, dropping up to 42.66% at beam size 1. Full remapping remains substantially worse than the preservation policy under warm-start training, indicating that short-budget continued training alone cannot repair topology disruption. (3) The SID topology-preserving principle makes even the simplest warm-start strategy highly effective: fixing old-item SIDs, inserting new-items, and warm-starting from the previous model improves old-item HitRate by 10% and new-item HitRate by 109%, validated across two non-overlapping update cycles. These findings distill into a concise deployment principle: insert new-items, freeze old-items; warm-start when possible, otherwise freeze; never change SIDs the model has already learned. A three-week online A/B test on TmallAPP shows UCTCVR +0.29% and GMV +1.16%.

Abstract: Language models can make personalized recommendations from rich item content, user history, and feedback. However, evaluating many candidates with full-text inputs incurs substantial online computation and limits serving throughput. We introduce DuoLM, a framework for fast and effective personalized recommendation that integrates a compact recommendation model with an efficient serving runtime. An encoder LLM learns compact item representations for offline reuse, while a personalized ranker LLM combines them with user history and feedback for recommendation. We train both models end to end with recommendation supervision and teacher distillation. The runtime further accelerates inference through virtual token IDs for embedding access, shared history computation, and CUDA Graph replay. Benchmark results show that our framework achieves recommendation quality comparable to that of the expert teacher model. On a single H100, it delivers approximately 26,400 candidate scores per second, about $60\times$ the throughput of the teacher served with vLLM and automatic prefix caching.

Abstract: Optimizing product listings for generative search requires language agents not only to rewrite effectively but also to continually self-improve by retaining successful strategies across diverse categories. However, existing approaches often treat structured listings as free-form text, leaving effective strategies implicit in prompts, search trajectories, or rewritten outputs rather than accumulating them as a reusable agentic memory. To address this limitation, we introduce GLOBE, an agentic framework that drives continual self-improvement by storing optimization knowledge in an evolving handbook of explicit, field-scoped behaviors. In each episode, an Editor applies relevant behaviors and receives rank-based feedback, which a Reflector converts into behavior updates. Crucially for dependable execution in real-world settings, a Moderator removes redundant proposals, a Curator commits accepted changes, and consistently harmful behaviors are deactivated. This cycle transforms episodic feedback into a persistent, field-scoped optimization memory. Experiments show that structured behaviors outperform monolithic prompt optimization, while feedback-driven evolution and moderation further improve effectiveness and reduce downside risk. The learned behaviors also remain effective on unseen product categories, suggesting that agentic memory evolution supports interpretable and transferable product optimization.

Abstract: Attribute-value extraction is important for building high-quality product catalogs in e-commerce. Existing systems either rely on generative or tagging strategies for extracting the attribute names and their corresponding values. Tagging models directly label spans in the source, making them fast and grounded, but they cannot produce values that require reformatting or generate rare attribute names outside their label space. Generative models avoid these drawbacks by generating both attribute names and values token by token, but this auto-regressive nature makes them slow and limits attribute coverage. We introduce hybrid generative-tagging decoding, a novel model architecture that combines tagging source spans and open-ended generation. For frequent attributes whose values are copyable, a tagging head extracts values directly from the source. This lets the generative head focus on the remainder - rare attributes outside the tagger's label space and non-copyable values, including implicit or normalized values. Integrated into vLLM inference, the hybrid decoder raises attribute coverage by 10% and extracts 63% more correct attribute-value pairs at no precision cost, while reducing latency by 44%.

Abstract: Recommender systems often need to balance multiple conflicting objectives. Controllable Multi-Objective Recommendation (CMOR) enables dynamic adjustment of objective trade-offs at inference time without retraining, adapting to shifting user or system requirements. Existing CMOR approaches typically rely on output-level reranking or parameter-generating modules, which either limit the scope of feasible trade-offs or incur substantial training and inference cost. Model merging provides a promising alternative by reusing multiple objective-specific models via parameter combination. However, its direct application to recommender systems is challenging due to the lack of a shared linear regime and cross-objective interference. To address these issues, we propose RecSoup, a lightweight two-stage approach for effective CMOR via specialized model merging. RecSoup first mitigates parameter interference across objectives through a pruning strategy that jointly considers objective-specific importance and cross-objective relevance. It then performs adaptive parameter merging via a learnable preference-to-coefficient mapping, allowing dynamic trade-off control at inference time without joint retraining or additional generative modules. Theoretically, we prove that RecSoup outperforms the parameter generation paradigm in effectiveness and efficiency, as it can fully navigate the Pareto front without smoothness constraints. Extensive experiments on multiple benchmarks demonstrate that RecSoup achieves improved Pareto front quality and effective controllability while delivers 165--228x faster inference-time latency than parameter generation approaches. Code is available at https://github.com/35798642/RECSOUP.git.

Abstract: We present TransX, a sequence-transduction framework and co-designed infrastructure for scaling industrial recommendation with explicit control over model-compute allocation and execution amortization. TransX separates sequence encoding, behavior memory, and candidate-conditioned fusion into independently scalable components. Training amortizes causal trajectory computation across behavioral prefixes and fans out memory states to candidates; serving disaggregates bulk prefill, nearline refresh, and online prediction through compact behavior-memory KV. We observe anisotropic scaling returns: encoder scaling may saturate beyond moderate depth at fixed memory resolution and fusion capacity; increasing supervision shifts the allocation frontier toward candidate-conditioned fusion; and the memory resolution needed to preserve quality remains roughly stable as model capacity grows. We train TransX on click and conversion, post-train its shared behavioral interface for retrieval and preference ranking, and explore generative recommendation offline. In separate production A/B experiments, the core TransX model improves CTR by 6.02% and conversion by 4.45%, while post-trained retrieval and preference-ranking variants improve CTR by 2.96% and conversion by 1.33%, respectively, over corresponding production baselines. Under matched training conditions, the training infrastructure increases supervised-target throughput by 84%. In serving, disaggregation reduces TransX's total serving cost by nearly 80% versus colocated serving at matched traffic and SLO, enabling approximately 5X the per-request model FLOPs of the HSTU-style production baseline at the same total serving cost. At matched per-request model FLOPs, TransX delivers approximately 4X higher request throughput per H200. These results show that online serving cost need not grow proportionally with recommendation model capacity.

Abstract: Group-based reinforcement learning enables large language model agents to learn from sparse outcome rewards in long-horizon tasks. Recent methods derive step-level credit by comparing observation-matched experiences across trajectories. However, identical observations can occur in different contexts, and matched peers are not equally informative for a given decision. Moreover, favorable terminal outcomes can obscure redundant actions or detours within successful trajectories. To address these two challenges, we propose Contextual Credit Policy Optimization (CCPO). For the first, CCPO weights observation-matched peers from other trajectories by contextual similarity, computed from frozen LLM representations and accumulated trajectory statistics, allowing peers from similar contexts to contribute more to a step's baseline. For the second, CCPO pairs historical credit, which compares the realized return against the weighted peer baseline, with future credit, which measures the change in contextual outcome value over a short realized continuation. A step is thus credited for the progress it makes, not only for its trajectory's eventual outcome. Both credits are computed in hindsight from rollouts already collected for training, enabling CCPO to address these challenges without an auxiliary value critic. Across two Qwen2.5-Instruct backbones (1.5B and 7B), CCPO achieves up to 97.4% success on ALFWorld and 79.4% on WebShop and outperforms GRPO by up to 21.6 and 18.5 points. Ablations show that contextual weighting outperforms uniform weighting, and that historical and future credit serve as complementary learning signals.

Abstract: Generative recommendation models have shown promising results by representing items as semantic IDs and generating these IDs autoregressively. We introduce an approach showing that explicitly incorporating hierarchical and attribute-level supervision into generative recommendation can simultaneously improve training efficiency and long-tail recommendation performance. Existing approaches typically optimize semantic ID prediction with a standard causal cross-entropy objective, which does not explicitly exploit the hierarchical structure of semantic IDs and can require substantial training to learn effective item representations. Moreover, data-sparse tail items often suffer from poor representations due to limited supervision, resulting in weaker long-tail recommendation performance. We address these challenges with HASTE (Hierarchical and Attribute-aware Supervision for generaTive rEcommendation), a unified training framework that introduces hierarchical and attribute-specific supervision. HASTE's Hierarchical Loss explicitly exploits the multi-level structure of semantic IDs, providing supervision at each level of the item hierarchy and substantially accelerating training convergence. Its Attribute-Specific Losses encourage the model to preserve fine-grained item attributes, leveraging the language model's inherent semantic knowledge to improve representations and recommendations for tail items. We extensively evaluate HASTE on a fashion recommendation benchmark and across multiple domains of the Amazon dataset. HASTE improves recommendation performance across datasets and item-frequency regimes. Notably, the Hierarchical Loss achieves approximately 2x faster convergence than the standard causal cross-entropy objective, substantially improving training efficiency while maintaining strong recommendation quality.

Abstract: Generative retrieval (GR) offers a promising paradigm for industrial search by unifying candidate generation within an autoregressive framework. However, its effectiveness is limited by two key challenges. First, existing Semantic ID (SID) tokenizers struggle to incorporate collaborative signals while maintaining effective codebook utilization, limiting their ability to distinguish items and achieve high-quality retrieval. Second, existing reinforcement learning (RL) methods suffer from severe reward sparsity under exact SID matching, leaving many sampled trajectories without effective learning signals and hindering preference optimization. To address these challenges, we propose WeGR, a unified generative retrieval framework consisting of two complementary components. Specifically, the Collaborative and Collapse-free Tokenizer (CoCo-VAE) incorporates query-to-item and item-to-item collaborative signals while mitigating codebook collapse through codebook refresh and complementary regularization, enabling more informative and balanced SID representations. Building upon these SIDs, Dynamic Horizon Policy Optimization (DHPO) progressively expands both reward evaluation and policy optimization horizons, reducing learning difficulty while leveraging behavioral reward signals at different levels of the SID hierarchy, without relying on auxiliary reward models. Extensive offline experiments demonstrate the effectiveness of WeGR in improving retrieval and ranking performance. Moreover, a two-month online A/B test in WeChat E-commerce Search demonstrates relative improvements of 2.539% in orders, 1.508% in click-through rate, and 1.381% in conversion rate, while reducing online CPU and GPU usage by 25.5% and 26.1%, respectively. The system has been successfully deployed across multiple search scenarios in WeChat, serving billions of users, generating tens of millions of PVs daily.

Abstract: Relevance prediction between user queries and items is a fundamental component of e-commerce search systems. Traditionally, high-quality relevance labels are obtained through human judgment, which is reliable but expensive, slow, and difficult to scale. Recent advances in large language models (LLMs) enable automated relevance judgment at scale, achieving performance comparable to human annotators. However, such systems often rely on large commercial LLMs with complex prompts, resulting in high inference costs and limited throughput. In this work, we propose a framework that bridges this gap by distilling a high-performing LLM-based relevance judgment system into efficient multimodal student models. We first curate a small, high-quality human-annotated dataset to serve as ground truth. Using this dataset, we develop a reliable LLM-as-a-judge pipeline through careful prompt engineering and iterative refinement. This pipeline is then used to generate large-scale relevance labels, which serve as training data for compact multimodal models. Our distilled models achieve performance close to the teacher system while significantly improving latency, cost, and scalability. This approach enables practical deployment of high-quality relevance prediction systems in production environments.

Abstract: E-commerce user preferences evolve dynamically within sessions, exhibiting concurrent and multi-centric interest distributions. While existing multimodal sequential recommendation methods leverage global attention to compress historical interactions, they inadequately capture the concurrent nature of multi-interest behaviors and fail to model the underlying multi-centric distribution of user intent. To tackle this issue, we propose OT-FlowRec, an optimal transport-guided multi-bridge flow matching framework that explicitly models multi-interest distributions to facilitate multimodal sequential recommendation. Unlike single-path generative models, OT-FlowRec constructs a mixture of parallel, interest-specific transport flows to model the multi-centric distribution characteristics, where each flow preserves a distinct user interest cluster via an optimal transport plan. Then, a geometric transport mechanism is designed to establish minimum-cost paths from the latent multi-interest distribution to target preference prototypes to provide multi-interest awareness. Moreover, a progressive time-sampling warmup strategy dynamically biases optimization weights toward the final states of the generation process, bridging the gap between continuous global distribution fitting and practical top-$k$ ranking metrics. This approach achieves explicit interest disentanglement, ensuring the preservation of a multi-peaked preference representation instead of collapsing it into a unimodal average. Extensive experiments on three public datasets demonstrate that OT-FlowRec outperforms existing state-of-the-art baselines.

Abstract: Structured knowledge retrieval acquires task-relevant evidence from structured knowledge sources for downstream tasks. Existing methods rely on unconstrained retrieval or predefined query-rule supervision, such as author-paper rules for scientific discovery and anatomy-drug rules for precision medicine. Such supervision is rarely available in practice. We propose Co-Adaptive Rule Routing and Constraining Retrieval (CoRE), which jointly learns which reasoning rule to assign to each query and how to retrieve evidence under that rule, without predefined query-rule pairs. A routing agent selects a rule, and its corresponding retriever traverses the knowledge base and ranks evidence that satisfies the rule's constraints. Because routing and retrieval are interdependent, CoRE alternates between reinforcement learning for the routing policy and contrastive learning for the rule-constrained retriever. The selected rules provide supervision for retriever training, while retrieval performance rewards the routing policy. This closed-loop process enables both components to co-adapt without explicit query-rule supervision. Experiments on STaRK-QA datasets across scientific discovery, biomedical reasoning, and e-commerce knowledge management demonstrate CoRE's effectiveness. Ablations show the benefit of co-adapting the router and retriever. Generation-based routing transfers across datasets and benefits from multi-domain pretraining followed by target-specific fine-tuning. Analysis of rule usage across queries also informs automatic rule refinement and discovery.

Abstract: Visual instance retrieval often fails when the same object appears at different apparent sizes in the query and gallery. We show that the dominant cause is usually not resolution loss but object-to-image (O2I) ratio mismatch: the object occupies different fractions of the two images. On a controlled benchmark of 3,021 Objaverse objects rendered at five camera distances, more than 80% of the cross-distance degradation is attributable to O2I mismatch rather than resolution for 9 of 12 pretrained backbones; multi-scale architectures cut the resolution-only effect to single digits yet remain equally susceptible. The failure is also asymmetric: tight queries retrieve more reliably against wide gallery images than the reverse. Guided by this analysis, query-side scale augmentation and an OWLv2 crop reranker reach state of the art on ILIAS 100M (29.2 mAP@1000 before reranking, 42.0 after) without training or modifying the precomputed gallery index, and a LoRA fine-tune matches the query-side gains at a single forward pass, showing that O2I robustness is learnable.

Abstract: Recommender systems leveraging generative models often generate item identifiers directly, rather than ranking catalog items by a recommendation score. Recent work extends beyond pure sequential interaction signals by incorporating item content and structured relationships among items, with two distinct directions emerging. Semantic IDs (SIDs) enrich item representations by replacing opaque, randomly initialized embeddings with hierarchically quantized discrete codes derived from item content. Knowledge-graph (KG) path reasoning instead generates entity-relation paths that ground recommendations in structured relationships between items, attributes, and external entities, thereby enriching the relational context. These two lines have complementary limitations: SID-based models lack relational grounding, while KG-based generative recommenders still represent items as arbitrary, opaque tokens tied to large embedding tables, limiting parameter sharing and generalization. We propose SPRIG, a generative recommender that integrates content-derived SIDs into KG path reasoning. SPRIG is trained on information-rich KG paths that terminate in items represented as discrete, content-derived tokens, combining the advantages of both approaches. We evaluate SPRIG on movie and music recommendation datasets against baselines spanning sequential language models, KG-augmented methods, and SID-based approaches. Our results show that SPRIG achieves competitive performance over prior generative models while using fewer parameters and a lower compute cost. Code: https://github.com/justinhangoebl/semantic-id-knowledge-graph-recommender

Abstract: As LLM agents decide on users' behalf which product to buy, which hotel to book, or which paper to cite, a preference for items from certain sources (the sites or services they come from) shapes what users receive and which sources are selected. We study source preference in end-to-end search with 12 agent models across three domains. Comparing items from different sources that satisfy the same requirements at the same position, we find that each model prefers some sources and avoids others in every domain, largely agreeing on which. This preference can outweigh how well items satisfy the request: an item satisfying one requirement fewer is selected about two-thirds of the time when it comes from a preferred source and the better one from a dispreferred source, but almost never in the reverse case. The information identifying an item's source affects selection by itself: hiding it weakens the preference, and relabeling an item with a preferred source raises its selection rate. We test two routes to this preference: training that rewards better items can make a source a shortcut for requirement satisfaction, and missing information can trigger preconceptions about the source. Supplying missing information or a prompt countering these preconceptions reduces source preference.

Abstract: In instruction-guided generative recommendation, LLM-based recommenders need to balance two goals: responding to the user's current request and aligning with the preferences in their interaction history. When the two conflict, history events can override the request. We show that turning the effect of individual history events into supervision faces two obstacles. First, the events that most influence a recommendation are not necessarily the ones that support the target item. Second, removing a misleading event can raise the target's score but a competing item's score even more, so a higher target score alone does not guarantee a better ranking. We propose Asymmetric Intervention-Guided Margin Supervision (AIMS), which converts the effect of removing individual history events into ranking supervision. For training requests already ranked correctly, a frozen reference model identifies request-specific deletions that improve both the target's score and its margin over a competitor near the recommendation cutoff. These margins serve as training targets, while the complete history is retained as input. Training combines cross-entropy with an asymmetric auxiliary loss that penalizes margin shortfalls and routes its gradient only through the competitor score. Inference is unchanged, requiring no history editing or deletion search. Across six LLM backbones on an industrial dataset and two public benchmarks, AIMS improves Recall and NDCG over strong baselines. Ablations support request-specific margins and asymmetric supervision, and the selected deletions preferentially remove constraint-violating history.

Abstract: Content-generation agents continuously receive impressions, clicks, conversions, and negative feedback from recommendation systems, providing real-world outcome signals for memory evolution. However, these signals are delayed and noisy, confounded by audience composition, placement, and recommendation policies, and may result from the combined influence of multiple memories, making accurate attribution difficult. Existing methods rely primarily on immediate feedback or semantic retrieval and therefore struggle to reliably translate recommendation outcomes into memory fitness. To address this challenge, we propose TIDE (Trajectory-Informed Directed Memory Evolution), an external memory evolution framework driven by delayed recommendation feedback. We further introduce Memory Evolution Gain (MEG), which measures the utility improvement of evolved memory over a no memory baseline on strictly future tasks. TIDE treats memory as a capacity-constrained population of experiences: temporal and semantic credit assignment estimates contextual fitness, while responsibility credit distributes outcome signals according to the memories referenced during generation. These signals are then used to reinforce, crossover, mutate, or evict memories. On an e-commerce membership marketing content-generation agent, TIDE achieves a +7.75-percentage-point MEG in offline temporal replay and significantly improves both unique click-through rate (UCTR) and activation rate in an online A/B test. On a delayed-label benchmark, TIDE achieves the lowest mean absolute error (MAE) and root mean squared error (RMSE) and the highest MEG among the compared methods, demonstrating its effectiveness.

Abstract: Industrial recommendation ranking models typically scale along two modeling axes: feature interaction over heterogeneous user, item, context, and cross features, and sequence modeling over long, informative, and multi-type user behavior histories. We find that scaling either capability in isolation is insufficient, as each exhibits a limited scaling ceiling and a suboptimal scaling-law slope. We conjecture that achieving a more favorable scaling-law slope requires jointly scaling both axes. To support this, we present HELIX, a purified and unified architecture for large-scale recommendation. HELIX interleaves sequence retrieval and feature interaction while enforcing one-way information flow from reusable sequence states to candidate-conditioned mix-tokens. This design preserves cross-depth communication between the two modeling axes while keeping user-side sequence computation amortizable, enabling flexible and asymmetric scaling of sequence modeling and feature interaction. Deployed in TikTok's e-commerce recommendation system, HELIX consistently improves offline CTR AUC, CVR AUC, and other ranking metrics. In online A/B tests, it achieves an approximately 6% increase in e-commerce video GMV per user.

Abstract: Agents need calibration when deployment conditions change: replacing a driving model, including a foundation-to-post-trained transition; crossing jurisdictions; or scaling across heterogeneous markets and sources. Interface compatibility alone does not establish capability retention or target-contract satisfaction. We formulate agent calibration as constrained behavioral adaptation across three interacting layers: information preservation, harness adaptation, and user acceptance; the layers apply to every scenario, not one-to-one to the three. The basic objective is non-degradation on prespecified capability measures while satisfying target requirements; aggregate improvement is stronger. Information calibration preserves independently validated source content still applicable to the target task. Harness calibration aligns observable artifacts at semantic checkpoints and repairs them through iteration, tool substitution, or local replanning within explicit budgets. User calibration enforces recipient-specific output contracts: templates, schemas, and section-level preferences. A global e-commerce example shows how shared standards coexist with site- and market-specific adapters and validation. We distinguish trainable policies from frozen-backbone configuration or controller optimization, and evidence verification from relative judgment and DPO/GRPO optimization. Recent harness-transfer and judge-validity studies motivate target-native execution records, separate audits of task validity and near-tie ranking, and matched target-native optimization controls. We propose held-out evaluations for model changes, cross-border adaptation, and scale, including a factorial test of source evidence and checkpoint repair and group-level reporting to prevent aggregate gains from masking local failures. This is a methodological proposal; implementation and empirical validation remain future work.

Abstract: Product attribute value extraction (AVE) is a fundamental task in e-commerce, aiming to identify specific values of predefined attributes from multimodal product profiles such as text and images. While multimodal large language models (MLLMs) have shown promise for AVE, they face challenges in extracting implicit attributes that require joint reasoning over visual and textual cues, often confusing semantically similar values. However, existing methods often fail to resolve such ambiguities because the correct value often depends on subtle multimodal cues that are easy to miss or override. To address this challenge, we propose Correcting to Predict (C2P), a framework that treats attribute extraction as a correction process. Given an initial pseudo-value such as a retrieved candidate or placeholder, the model learns to correct it using multimodal evidence. During training, diverse pseudo-values help the model learn evidence-based correction behavior, and a self-consistency refinement stage further reduces sensitivity to pseudo-value perturbations. At inference, a fixed placeholder triggers the learned correction behavior, enabling efficient single-pass prediction without online retrieval or iterative refinement. We evaluate C2P on a public benchmark and a large-scale industrial dataset. Offline results show that C2P outperforms strong baselines, with notable gains on ambiguous attributes. Online A/B tests on AliExpress further show consistent improvements in seller adoption, attribute completeness, and user engagement, validating C2P's effectiveness and efficiency in real-world deployment.

Abstract: Sequential recommenders typically rely on learnable positional embeddings to encode the order of user interactions. In this work, we ask whether this ordinal signal can be replaced by a structural one derived from the item space. We propose to use Laplacian positional embeddings in SASRec: we build an item co-occurrence graph from training interactions, compute eigenvectors of its symmetric normalized Laplacian, and use them as frozen graph-derived positional embeddings. The backbone architecture and training objective remain unchanged. Experiments on four public sequential-recommendation benchmarks show that this simple replacement improves SASRec performance on most ranking metrics and remains competitive with strong positional and temporal encoding baselines. These findings indicate that item-item graph structure can be an effective substitute for standard ordinal positional embeddings in sequential recommendation.

Abstract: Open rerankers trained for general web retrieval transfer imperfectly to e-commerce, where ranking decisions depend not only on topical relevance but also on user preferences, product constraints, and comparative product fit. These preference signals are difficult to supervise at scale: real search traffic provides authentic queries and candidates but no clean pairwise labels. We present ZooWork-ShopRanker, a family of e-commerce rerankers (0.6B, 4B, and 8B) aligned to judge-labeled shopping preference. Training pairs are labeled by a panel of reasoning large language models (LLMs) from different families acting as a preference oracle, with position-debiased judgments and agreement tiers, and the rerankers are trained on these labels. The aligned 8B flagship then serves as a distillation teacher for the efficient 4B and 0.6B models, which are fit to its scores and sharpened on judged pairs. To measure progress, we introduce ShopRank-Bench, a contamination-limited benchmark of ~10,000 private-traffic preference pairs in both text formats, tiered by how many judge families committed to each label. ZooWork-ShopRanker-8B and -4B significantly outperform the strongest open reranker baseline, every model significantly beats its own un-aligned base, and ZooWork-ShopRanker-0.6B beats its size peer; the gains hold in both formats and extend to common MTEB benchmarks. We release the models and the dual-format ShopRank-Bench to facilitate further research.

Abstract: Semantic IDs (SIDs) encode each catalog item as a short token sequence, enabling generative recommenders to predict the next item autoregressively. Reasoning-enhanced variants, an increasingly common extension, first generate a textual trace and then decode a next-item SID by beam search. Such recommenders are commonly trained with group-relative policy optimization under an exact-match SID reward, which is sparse in large catalogs. Two failure modes follow. When all rollouts in a group miss the target, the group yields zero advantage and no learning signal. Rollouts sharing the same SID reward receive identical advantages, however much their traces differ. In both cases the reward reflects only the decoded SID, never the reasoning that produced it. This creates a credit-assignment gap. We address this gap with retrieval-grounded query attribution. Each trace is structured into a history summary, a set of interest hypotheses, and a final SID. A frozen retriever executes every hypothesis as a catalog query, so that each hypothesis becomes independently verifiable rather than judged only through the final SID. A rollout is rewarded when any of its queries retrieves the target within the \mbox{top-$K$}, and per-query hit indicators localize that reward to individual hypotheses. Credit is thus assigned at the span level: only hypotheses that individually hit receive positive retrieval advantage, while the retrieval channel never updates the final SID span. Rollouts that share a SID reward can therefore receive different updates. Across experiments on three Amazon Reviews datasets, this yields consistent improvements in SID recommendation. On Video Games, an oracle analysis further reveals the potential of interest-conditioned SID decoding: selecting the target-relevant query among generated interests improves both recall and ranking.

Abstract: Generative recommendation reformulates item retrieval as sequence generation, allowing a unified model to directly generate the next item from a user's interaction history. Semantic IDs further make this paradigm effective and scalable by representing each item as discrete codes, enabling knowledge sharing among semantically related items. Recent studies introduce explicit reasoning before Semantic-ID generation, helping models summarize user interests and infer possible preference transitions. However, reasoning is not inherently beneficial: Inaccurate or uninformative reasoning may mislead subsequent item generation and ultimately degrade recommendation performance. This raises a central challenge: how can a recommender select and learn effective reasoning traces and progressively evolve toward better reasoning from its own generations? In this work, we propose Evo-Rec, a three-stage framework for learning better reasoning and further enhancing it through reinforcement learning. First, we align Semantic IDs with their textual and behavioral contexts, enabling the model to understand and generate item identifiers. Second, we sample multiple candidate reasoning traces and retain those that improve the prediction of the ground-truth item, providing a stronger reasoning initialization through supervised fine-tuning. Third, we further optimize the reasoning policy through reinforcement learning with catalog-constrained item generation and ranking-aware recommendation feedback. Experiments on three Amazon Review benchmarks show that Evo-Rec consistently outperforms discriminative, generative, and reasoning-enhanced recommenders across all evaluation metrics. These results demonstrate the effectiveness of our framework in learning better reasoning for SID-based generative recommendation.

Abstract: Visual search on large e-commerce catalogs must serve both "similarity" queries that ask for items resembling an uploaded image and "modifier" queries that comprise an image and text describing a desired modification (e.g. a color change or style swap). The latter is the setting known as composed image retrieval (CIR). Existing CIR methods, however, treat relevance as binary and train on triplets with a single positive target - a poor fit for real catalogs where many candidates partially satisfy a user query and ranking across that partial-match spectrum drives the customer experience. We propose a methodology for training CIR retrievers on graded relevance, consisting of: (i) a VLM to curate training data, generating both queries (object detection + modifier synthesis) and 4-level relevance labels without manual annotation, (ii) an iterative relevance-feedback loop that expands the training set by mining hard negatives from the in-training retriever, and (iii) a hierarchy-aware angular objective to train the retriever directly on the graded labels rather than collapsing them to a binary split. We call this methodology GradCIR and instantiate it on a PaliGemma2 bi-encoder trained on 3.5M graded pairs curated from raw Walmart catalog data. A controlled graded-vs-binary ablation isolates the supervision granularity and shows lift of 4.9%-5.9% in NDCG@10. The same recipe applied to other multimodal encoders lifts early-fusion backbones by up to 8.5% NDCG@10. On the public FashionIQ benchmark, GradCIR (applied to PaliGemma2) reaches 0.6703 average recall when fine-tuned, slightly ahead of the strongest peer-reviewed supervised baseline we compare against, and matching or exceeding all published CLIP-L-class zero-shot CIR methods. The system is deployed in production at Walmart, where it's serving live visual-search user traffic.

Abstract: In this reproducibility study, we investigate the transparency and scrutability of recommender systems enhanced by incorporating generated natural-language user profiles that represent user preferences. The original paper explores the synthesis of user profiles from raw user-generated review text across domains such as movies and accommodations (Amazon Movies & TV, TripAdvisor). Crucially, these natural-language user profiles enable direct user interaction and intervention, allowing users to customize recommendations by correcting misattributed preferences or addressing cold-start settings. We successfully reproduce the core findings of the original study. Additionally, we extend the evaluation by conducting systematic context ablation experiments, multi-seed stability across five distinct random seeds to establish statistical reliability, and a mechanistic interpretability analysis using the nnsight framework to probe internal model representations under counterfactual profile perturbations. Our findings verify the original paper's claim that User Profile Recommendation (UPR) achieves competitive performance under its test-set reranking protocol and makes recommendations more transparent. Perturbing the natural-language profiles does change predictions, but it shifts predicted ratings uniformly across genres with no detectable genre-selective effect, leaving rankings unchanged even under direct activation steering. We trace this back to the rating-regression objective rather than the profile interface, with ranking-objective models clearly exceeding in this task.

Abstract: The increasing demand for user privacy and compliance with regulations such as GDPR has made machine unlearning a fundamental requirement for modern recommender systems. However, Sequential Recommender Systems (SRS) pose unique challenges for unlearning due to their reliance on temporal interaction patterns. Existing approaches either require computationally prohibitive full retraining or fail to account for the sequential nature of user behavior. We propose SURF (Subtractive Updates for Recommender Forgetting), a lightweight framework for approximate machine unlearning in SRS. SURF operates in three stages: (i) identifying the neighborhood of the item to forget in the embedding space, (ii) training an auxiliary model on this compact local subset, and (iii) subtracting the auxiliary model's scores from the original model at inference time. Experiments against five baselines on 7 datasets show that SURF achieves unlearning effectiveness comparable to full retraining while substantially reducing computational cost, yielding up to a 32% improvement in NDCG@20 while requiring just 2% of the original retraining baseline time budget. We share our code at https://github.com/FilippoBetello/SURF.

Abstract: We propose a personalized capping framework (PCap) to improve the diversity in Facebook Marketplace by introducing user-level diversity constraints at the retrieval stage. PCap models individual diversity preferences using Shannon entropy-based scoring, segments users into diversity buckets, and applies personalized category caps during multi-source candidate retrieval. To navigate the high-dimensional parameter space of per-bucket caps, we leverage an automated online optimization method called Parameter Tuning Sequence. Large-scale online experiments demonstrate that PCap significantly improves users' browsing experience shown in engagement metrics. This work provides practical insights into integrating personalized diversity into industrial retrieval systems.

Abstract: Transformers have shown promising performance in LLMs due to their outstanding scalability, several studies have investigated the scalability of Transformers for industrial recommendation. They typically rely on a single ranking model to optimize both sparse and dense parameters from scratch, resulting in substantial computational resource consumption and slow convergence. Fortunately, the pre-training models offer an effective solution to the above issues by providing favorable initialization of both sparse and dense parameters for the subsequent ranking. However, they still face two major limitations: (1) Since the input features used in pre-training and ranking are usually inconsistent, directly transferring dense parameters from pre-training to ranking may lead to negative transfer. (2) Multi-epoch training during the ranking process may result in the overfitting of sparse parameters, while freezing the sparse parameters limits their adaptability to the ranking objectives. To this end, we propose a Scaling Transformer for Industrial Recommendation via Transferable Generative Pre-training, termed LazFormer. Specifically, we first present a generative pre-training module to autoregressively generate sequential features, providing favorable initialization of both sparse and dense parameters for the subsequent ranking. To solve the negative transfer of dense parameters, we propose a transferable residual adapter that injects additional ranking-specific features into ranking in a residual manner. Moreover, a request-aware ranking module integrates long-sequence compression, hybrid sparse attention, and a request-aware paradigm to efficiently model users' long sequences. Besides, we further propose an asymmetric multi-epoch training strategy that resets sparse parameters while continuously accumulating dense parameters across epochs, alleviating the overfitting of sparse parameters.

Abstract: Integrating recall and pre-ranking in e-commerce search requires candidate generation to account for relevance, personalization, and business value before final ranking. To this end, we present VARG, a generative retrieval system for Tmall App search that directly admits generated item candidates to the existing final ranker. VARG-ID constructs semantic prefixes using RQ-VAE, enhances search relevance through bidirectional query-item contrastive learning, and combines these prefixes with a value-ordered third token to provide fine-grained item addresses and a business-value prior. Three-stage supervised fine-tuning progressively learns item-to-identifier mappings, query-semantic retrieval, and personalized retrieval. Personalized model training combines value-aware and hierarchy-aligned supervision with expanded user context, and uses local ordinal supervision (LO-SFT) to learn the local within-cluster ordering encoded by the third token. Prefix-GRPO combines gated rewards based on output legality, user behavior, ranker advantage, and search relevance with prefix-aware token weighting to align candidate generation with business value and ranking objectives. Coordinated daily product and model updates preserve existing item addresses while incorporating new products and behavioral feedback. Offline experiments on tens of millions of products validate identifier stability and demonstrate gains in retrieval quality and head-level value recall from SFT strategies and Prefix-GRPO over their respective baselines. In a 14-day online A/B test covering 20% of search traffic, VARG directly admits generated candidates to the final ranker and improves GMV by 1.45%, per-user IPV by 0.22%, and PCTR by 0.31%. Online shopping-guide query evaluations further show that VARG maintains competitive relevance with a smaller candidate quota.

Abstract: Modeling ultra-long user behavior sequences is crucial for industrial recommendation and online advertising, yet directly feeding thousands of historical actions into ranking models is computationally prohibitive, while truncation discards long-range signals. Existing lifelong-interest methods retrieve target-relevant behaviors for each candidate, coupling long-sequence modeling with candidate scoring and repeated online cost. Recent target-independent compression methods enable cached user summaries, but often append query tokens at the sequence end and use bidirectional encoding, producing unordered and redundant summaries that overlook temporal structure. We propose ChronicleRec, a pre-train-and-transfer framework that compresses an ultra-long behavior sequence once into a chronologically ordered set of Chronicle Tokens. ChronicleRec applies a recency-aware multi-granularity merge, preserving recent behaviors while coarsening distant history. It then interleaves query tokens with the merged sequence and uses a causal encoder, so each query summarizes only the history before its temporal anchor. A multi-horizon design masks different recent-history windows across parallel branches to learn complementary long-range interests. The compressor is pre-trained with a mask-and-predict objective that reconstructs held-out recent behaviors from compressed older history, aligning historical signals with near-present intent. Since Chronicle Tokens are target-independent, they can be cached per user, decoupling ultra-long sequence modeling from online candidate scoring. Experiments on KuaiRand and Tencent AdLive show that ChronicleRec outperforms recent-window and single-pass compression baselines while approaching full-attention performance. Token analyses reveal temporally organized and complementary representations, and a seven-day online A/B test confirms significant production gains.

Abstract: Customers rely on specific product attributes to compare products and make purchasing decisions, but e-commerce catalogs are messy and unstructured, making it difficult to identify which attributes matter most and extract them at scale. Standard Attribute Value Extraction (AVE) systems treat all attributes equally, producing large, inconsistent attribute sets that do not reflect the factors consumers use to differentiate products. We introduce a two-stage LLM pipeline that first discovers a compact, ranked schema of purchase-discriminative attributes for each product category, then extracts their values from catalog text using a fine-tuned compact LLM (Qwen3-4B) with Hyper-Parallel Decoding (HPD). This pipeline achieves 85% extraction accuracy, on par with the foundational LLM it was distilled from, while reducing inference costs by 92% over foundational LLMs, enabling production-scale use for product discovery and catalog enrichment. The resulting category-level structured representations effectively constitute automatically constructed product knowledge bases, providing consistent, comparable attributes across varied product categories that can ground downstream knowledge-intensive applications.

Abstract: E-commerce platforms increasingly display clickable query suggestions alongside items in the user feed, enabling users to refine or expand their intent without manually reformulating queries. Existing approaches either mine suggestions from historical logs -- limited to past behavior and blind to long-tail, personalized intents -- or rely on off-the-shelf LLMs whose lack of platform-specific knowledge yields fluent but generic queries disconnected from real click behavior. We propose EAGER (Enrich-and-AliGn gEnerative Query Recommendation), a two-stage framework for generating query suggestions from clicked items. In the enrichment stage, supervised fine-tuning (SFT) follows a four-stage curriculum that scales information richness (from item-only to user-conditioned) and reasoning depth (from direct to chain-of-thought). Each stage incorporates rationale augmentation, diversity regularization, and self-distillation. In the alignment stage, we post-train via GRPO with a hybrid reward of multiple rule-based business signals and a preference-aware click reward. Extensive offline experiments and online A/B test demonstrate the effectiveness of EAGER, which has been deployed in production at a major e-commerce platform.