arXiv subjects

Hangyu Wang

Publications and source records attributed to Hangyu Wang.

10 recordsLinked to original sources

arXiv2026

SequenceO1: End-to-End Ultra-Long (100K) Sequence Modeling in Recommendation with Low-Rank Caching

Modeling long-term user behavior is central to sequential recommendation and billion-scale industrial recommender systems, yet production ranking models operate under strict latency, memory, communication, and training-throughput constraints. At the 100K scale, the challenge extends beyond attention complexity: raw sequence features must be stored, transferred, and repeatedly processed during training and online serving. Existing approaches based on history truncation, multi-stage behavior retrieval, compressed lifelong histories, or train-short/infer-long extrapolation either weaken end-to-end optimization or retain substantial length-dependent cost. We present SequenceO1, an end-to-end framework for ultra-long user behavior sequence modeling, deployed at full traffic on Douyin with histories of up to 100K interactions. SequenceO1 follows a compress-then-reason design. Its Sketch Attention (SA) uses learnable prototypes and prototype-wise normalization to compress the raw history into a fixed-size, target-agnostic user representation. Target-conditioned Stacked Target-to-History Cross Attention (STCA) then models complementary time scales: a recent 10K suffix for short-term interests and the compact sketch for long-term preferences. To make training and inference practical, SequenceO1 combines low-rank user representation caching, multi-request user-level batching, pipeline lift, and a fused FlashSA kernel to amortize feature storage, communication, and computation across targets, training instances, and consecutive requests. Production experiments show consistent offline and online gains, while the compact cached sketch retains most of the benefit of directly scaling end-to-end sequence ranking to 100K. These results provide a practical model-system approach to efficient attention, sequence compression, and scalable long-sequence and long-context recommendation systems.

cs.IR↗

arXiv2026

PlexRL: Cluster-Level Orchestration of Serviceized LLM Execution for RLVR

Reinforcement learning with verifiable rewards (RLVR) has recently unlocked strong reasoning capabilities in large language models (LLMs), triggering rapid exploration of new algorithms and data. However, RLVR training is notoriously inefficient: long-tailed rollouts, tool-induced stalls, and asymmetric resource requirements between rollout and training introduce substantial idle time that cannot be eliminated by job-local optimizations such as synchronous pipelining, asynchronous rollout, or colocated execution. We argue that this inefficiency is structural. While idle gaps are unavoidable within individual RLVR jobs, they are largely anti-correlated across jobs and therefore exploitable at the cluster level. Leveraging this observation, we present PlexRL, a cluster-level runtime for multiplexing unified LLM services across RLVR jobs. By centrally managing model placement, state transitions, and function-level scheduling under strict affinity constraints, PlexRL time-slices LLM execution across jobs to fill otherwise idle periods without expensive model migration. Our implementation and evaluations demonstrate that PlexRL significantly improves effective cluster capacity and reduces user GPU hour cost by maximum 37.58% while preserving algorithmic flexibility and introducing minimal per-job overhead.

cs.DC↗

arXiv2025

Make It Long, Keep It Fast: End-to-End 10K Long User Behavior Sequence Modeling for Billion-Scale Douyin Recommendation

Short-video recommenders such as Douyin must exploit extremely long user behavior histories without breaking latency or cost budgets. We present an end-to-end industrial recommender system that scales long-sequence recommendation modeling to 10K-length histories in production. First, we introduce Stacked Target-to-History Cross Attention (STCA), which replaces history self-attention with stacked cross-attention from the target to the history, reducing complexity from quadratic to linear in sequence length and enabling efficient end-to-end training over long user behavior sequences. Second, we propose Request Level Batching (RLB), a user-centric batching scheme that aggregates multiple targets for the same user/request to share the user-side encoding, substantially lowering sequence-related storage, communication, and compute without changing the learning objective. Third, we design a length-extrapolative training strategy -- train on shorter windows, infer on much longer ones -- so the model generalizes to 10K-scale histories without additional training cost. Across offline and online experiments, we observe predictable, monotonic gains as we scale history length and model capacity, mirroring the scaling law behavior observed in large language models. Deployed at full traffic on Douyin, our system delivers significant improvements on key engagement metrics while meeting production latency, demonstrating a practical path to scaling end-to-end ultra-long sequence recommendation to the 10K regime.

cs.LG↗

arXiv2025

RankMixer: Scaling Up Ranking Models in Industrial Recommenders

Recent progress on large language models (LLMs) has spurred interest in scaling up recommendation systems, yet two practical obstacles remain. First, training and serving cost on industrial Recommenders must respect strict latency bounds and high QPS demands. Second, most human-designed feature-crossing modules in ranking models were inherited from the CPU era and fail to exploit modern GPUs, resulting in low Model Flops Utilization (MFU) and poor scalability. We introduce RankMixer, a hardware-aware model design tailored towards a unified and scalable feature-interaction architecture. RankMixer retains the transformer's high parallelism while replacing quadratic self-attention with multi-head token mixing module for higher efficiency. Besides, RankMixer maintains both the modeling for distinct feature subspaces and cross-feature-space interactions with Per-token FFNs. We further extend it to one billion parameters with a Sparse-MoE variant for higher ROI. A dynamic routing strategy is adapted to address the inadequacy and imbalance of experts training. Experiments show RankMixer's superior scaling abilities on a trillion-scale production dataset. By replacing previously diverse handcrafted low-MFU modules with RankMixer, we boost the model MFU from 4.5\% to 45\%, and scale our ranking model parameters by 100x while maintaining roughly the same inference latency. We verify RankMixer's universality with online A/B tests across two core application scenarios (Recommendation and Advertisement). Finally, we launch 1B Dense-Parameters RankMixer for full traffic serving without increasing the serving cost, which improves user active days by 0.3\% and total in-app usage duration by 1.08\%.

cs.IR↗

arXiv2024

Efficiency Unleashed: Inference Acceleration for LLM-based Recommender Systems with Speculative Decoding

The past few years have witnessed a growing interest in LLM-based recommender systems (RSs), although their industrial deployment remains in a preliminary stage. Most existing deployments leverage LLMs offline as feature enhancers, generating augmented knowledge for downstream tasks. However, in recommendation scenarios with numerous users and items, even offline knowledge generation with LLMs demands significant time and computational resources. This inefficiency arises from the autoregressive nature of LLMs. A promising solution is speculative decoding, a Draft-Then-Verify approach that increases the number of tokens generated per decoding step. In this work, we first identify recommendation knowledge generation as a highly fitting use case for retrieval-based speculative decoding. Then, we discern its two characteristics: (1) the vast number of items and users in RSs leads to retrieval inefficiency, and (2) RSs exhibit high diversity tolerance for LLM-generated text. Building on these insights, we introduce Lossless Acceleration via Speculative Decoding for LLM-based Recommender Systems (LASER), which features a Customized Retrieval Pool to enhance retrieval efficiency and Relaxed Verification to improve the acceptance rate of draft tokens. LASER achieves a 3-5x speedup on public datasets and saves about 67\% of computational resources during the online A/B test on a large-scale advertising scenario with lossless downstream recommendation performance. Our code is available at https://github.com/YunjiaXi/LASER

cs.IR↗

arXiv2024

DisCo: Towards Harmonious Disentanglement and Collaboration between Tabular and Semantic Space for Recommendation

Recommender systems play important roles in various applications such as e-commerce, social media, etc. Conventional recommendation methods usually model the collaborative signals within the tabular representation space. Despite the personalization modeling and the efficiency, the latent semantic dependencies are omitted. Methods that introduce semantics into recommendation then emerge, injecting knowledge from the semantic representation space where the general language understanding are compressed. However, existing semantic-enhanced recommendation methods focus on aligning the two spaces, during which the representations of the two spaces tend to get close while the unique patterns are discarded and not well explored. In this paper, we propose DisCo to Disentangle the unique patterns from the two representation spaces and Collaborate the two spaces for recommendation enhancement, where both the specificity and the consistency of the two spaces are captured. Concretely, we propose 1) a dual-side attentive network to capture the intra-domain patterns and the inter-domain patterns, 2) a sufficiency constraint to preserve the task-relevant information of each representation space and filter out the noise, and 3) a disentanglement constraint to avoid the model from discarding the unique information. These modules strike a balance between disentanglement and collaboration of the two representation spaces to produce informative pattern vectors, which could serve as extra features and be appended to arbitrary recommendation backbones for enhancement. Experiment results validate the superiority of our method against different models and the compatibility of DisCo over different backbones. Various ablation studies and efficiency analysis are also conducted to justify each model component.

cs.IR↗

arXiv2024

Towards Efficient and Effective Unlearning of Large Language Models for Recommendation

The significant advancements in large language models (LLMs) give rise to a promising research direction, i.e., leveraging LLMs as recommenders (LLMRec). The efficacy of LLMRec arises from the open-world knowledge and reasoning capabilities inherent in LLMs. LLMRec acquires the recommendation capabilities through instruction tuning based on user interaction data. However, in order to protect user privacy and optimize utility, it is also crucial for LLMRec to intentionally forget specific user data, which is generally referred to as recommendation unlearning. In the era of LLMs, recommendation unlearning poses new challenges for LLMRec in terms of \textit{inefficiency} and \textit{ineffectiveness}. Existing unlearning methods require updating billions of parameters in LLMRec, which is costly and time-consuming. Besides, they always impact the model utility during the unlearning process. To this end, we propose \textbf{E2URec}, the first \underline{E}fficient and \underline{E}ffective \underline{U}nlearning method for LLM\underline{Rec}. Our proposed E2URec enhances the unlearning efficiency by updating only a few additional LoRA parameters, and improves the unlearning effectiveness by employing a teacher-student framework, where we maintain multiple teacher networks to guide the unlearning process. Extensive experiments show that E2URec outperforms state-of-the-art baselines on two real-world datasets. Specifically, E2URec can efficiently forget specific data without affecting recommendation performance. The source code is at \url{https://github.com/justarter/E2URec}.

cs.IR↗

arXiv2023

FLIP: Fine-grained Alignment between ID-based Models and Pretrained Language Models for CTR Prediction

Click-through rate (CTR) prediction plays as a core function module in various personalized online services. The traditional ID-based models for CTR prediction take as inputs the one-hot encoded ID features of tabular modality, which capture the collaborative signals via feature interaction modeling. But the one-hot encoding discards the semantic information included in the textual features. Recently, the emergence of Pretrained Language Models(PLMs) has given rise to another paradigm, which takes as inputs the sentences of textual modality obtained by hard prompt templates and adopts PLMs to extract the semantic knowledge. However, PLMs often face challenges in capturing field-wise collaborative signals and distinguishing features with subtle textual differences. In this paper, to leverage the benefits of both paradigms and meanwhile overcome their limitations, we propose to conduct Fine-grained feature-level ALignment between ID-based Models and Pretrained Language Models(FLIP) for CTR prediction. Unlike most methods that solely rely on global views through instance-level contrastive learning, we design a novel jointly masked tabular/language modeling task to learn fine-grained alignment between tabular IDs and word tokens. Specifically, the masked data of one modality (IDs and tokens) has to be recovered with the help of the other modality, which establishes the feature-level interaction and alignment via sufficient mutual information extraction between dual modalities. Moreover, we propose to jointly finetune the ID-based model and PLM by adaptively combining the output of both models, thus achieving superior performance in downstream CTR prediction tasks. Extensive experiments on three real-world datasets demonstrate that FLIP outperforms SOTA baselines, and is highly compatible with various ID-based models and PLMs. The code is at \url{https://github.com/justarter/FLIP}.

cs.IR↗

arXiv2023

ClickPrompt: CTR Models are Strong Prompt Generators for Adapting Language Models to CTR Prediction

Click-through rate (CTR) prediction has become increasingly indispensable for various Internet applications. Traditional CTR models convert the multi-field categorical data into ID features via one-hot encoding, and extract the collaborative signals among features. Such a paradigm suffers from the problem of semantic information loss. Another line of research explores the potential of pretrained language models (PLMs) for CTR prediction by converting input data into textual sentences through hard prompt templates. Although semantic signals are preserved, they generally fail to capture the collaborative information (e.g., feature interactions, pure ID features), not to mention the unacceptable inference overhead brought by the huge model size. In this paper, we aim to model both the semantic knowledge and collaborative knowledge for accurate CTR estimation, and meanwhile address the inference inefficiency issue. To benefit from both worlds and close their gaps, we propose a novel model-agnostic framework (i.e., ClickPrompt), where we incorporate CTR models to generate interaction-aware soft prompts for PLMs. We design a prompt-augmented masked language modeling (PA-MLM) pretraining task, where PLM has to recover the masked tokens based on the language context, as well as the soft prompts generated by CTR model. The collaborative and semantic knowledge from ID and textual features would be explicitly aligned and interacted via the prompt interface. Then, we can either tune the CTR model with PLM for superior performance, or solely tune the CTR model without PLM for inference efficiency. Experiments on four real-world datasets validate the effectiveness of ClickPrompt compared with existing baselines.

cs.IR↗

arXiv2023

GMOCAT: A Graph-Enhanced Multi-Objective Method for Computerized Adaptive Testing

Computerized Adaptive Testing(CAT) refers to an online system that adaptively selects the best-suited question for students with various abilities based on their historical response records. Most CAT methods only focus on the quality objective of predicting the student ability accurately, but neglect concept diversity or question exposure control, which are important considerations in ensuring the performance and validity of CAT. Besides, the students' response records contain valuable relational information between questions and knowledge concepts. The previous methods ignore this relational information, resulting in the selection of sub-optimal test questions. To address these challenges, we propose a Graph-Enhanced Multi-Objective method for CAT (GMOCAT). Firstly, three objectives, namely quality, diversity and novelty, are introduced into the Scalarized Multi-Objective Reinforcement Learning framework of CAT, which respectively correspond to improving the prediction accuracy, increasing the concept diversity and reducing the question exposure. We use an Actor-Critic Recommender to select questions and optimize three objectives simultaneously by the scalarization function. Secondly, we utilize the graph neural network to learn relation-aware embeddings of questions and concepts. These embeddings are able to aggregate neighborhood information in the relation graphs between questions and concepts. We conduct experiments on three real-world educational datasets, and show that GMOCAT not only outperforms the state-of-the-art methods in the ability prediction, but also achieve superior performance in improving the concept diversity and alleviating the question exposure. Our code is available at https://github.com/justarter/GMOCAT.

cs.IR↗