Large language models (LLMs) have become an important semantic infrastructure for modern recommender systems. A prevailing paradigm integrates LLM-derived semantic embeddings with collaborative representations via representation alignment, implicitly assuming that the two views encode a shared latent entity and that stronger alignment yields better results. We formalize this assumption as the global low-complexity alignment hypothesis and argue that it is stronger than necessary and often structurally mismatched with real-world recommendation settings. We propose a complementary perspective in which semantic and collaborative representations are treated as partially shared yet fundamentally heterogeneous views, each containing both shared and view-specific factors. Under this shared-plus-private latent structure, enforcing global geometric alignment may distort local structure, suppress view-specific signals, and reduce informational diversity. To support this perspective, we develop complementarity-aware diagnostics that quantify overlap, unique-hit contribution, and theoretical fusion upper bounds. Empirical analyses on sparse recommendation benchmarks reveal low item-level agreement between semantic and collaborative views and substantial oracle fusion gains, indicating strong complementarity. Furthermore, controlled alignment probes show that low-capacity mappings capture only shared components and fail to recover full collaborative geometry, especially under distribution shift. These findings suggest that alignment should not be treated as the default integration principle. We advocate a shift from alignment-centric modeling to complementarity fusion-centric, complementarity-aware design, where shared factors are selectively integrated while private signals are preserved. This reframing provides a principled foundation for the next generation of LLM-enhanced recommender systems.

The escalating complexity of industrial recommendation systems, characterized by diverse user behaviors and cross-domain application scenarios, necessitates advanced multi-task and multi-domain learning paradigms. Existing methods often struggle with efficient knowledge transfer across tasks and domains due to semantic gaps and distribution shifts. To address these challenges, we propose the Perception-Aware Multi-Task Transformer Network for Personalized Multi-Domain Adaptation (PMTA), a unified framework that integrates three key innovations: First, the Task Prompt Encoding (TPE) module dynamically generates prompts by synthesizing personalized user data with task-specific information. Second, the Transformer-based Multi-Task Perception (TMPN) network enables adaptive cross-task knowledge transfer through attention mechanisms. Third, the Multi-Domain Adaptation (MDAN) component captures domain-specific behavior patterns via learnable prior information. Experimental results demonstrate PMTA's effectiveness, achieving 0.168% increase in watch time and significant improvements in engagement metrics (AAD: +0.0113%, AAH: +0.0608%). Deployed on Douyin and Douyin Lite, it significantly improves recommendation quality and drives commercial success.

Sequential recommendation, a critical task in recommendation systems, predicts the next user action based on the understanding of the user's historical behaviors. Conventional studies mainly focus on cross-behavior modeling with self-attention based methods while neglecting comprehensive user interest modeling for more dimensions. In this study, we propose a novel sequential recommendation model, Pyramid Mixer, which leverages the MLP-Mixer architecture to achieve efficient and complete modeling of user interests. Our method learns comprehensive user interests via cross-behavior and cross-feature user sequence modeling. The mixer layers are stacked in a pyramid way for cross-period user temporal interest learning. Through extensive offline and online experiments, we demonstrate the effectiveness and efficiency of our method, and we obtain a +0.106% improvement in user stay duration and a +0.0113% increase in user active days in the online A/B test. The Pyramid Mixer has been successfully deployed on the industrial platform, demonstrating its scalability and impact in real-world applications.

In large-scale recommendation systems, modeling long-term user interests is progressively gaining attention among researchers and practitioners. Existing work, such as SIM and TWIN, typically employs a two-stage approach to model long-term user behavior sequences for efficiency concerns. The first stage rapidly retrieves a subset of sequences related to the target item from a long sequence using a search-based mechanism namely the General Search Unit (GSU), while the second stage calculates the interest scores using the Exact Search Unit (ESU) on the retrieved results. Given the extensive length of user behavior sequences spanning the entire life cycle, potentially reaching up to 10^6 in scale, there is currently no effective solution for fully modeling such expansive user interests. To overcome this issue, we introduced TWIN-V2, an enhancement of TWIN, where a divide-and-conquer approach is applied to compress life-cycle behaviors and uncover more accurate and diverse user interests. Specifically, a hierarchical clustering method groups items with similar characteristics in life-cycle behaviors into a single cluster during the offline phase. By limiting the size of clusters, we can compress behavior sequences well beyond the magnitude of 10^5 to a length manageable for online inference in GSU retrieval. Cluster-aware target attention extracts comprehensive and multi-faceted long-term interests of users, thereby making the final recommendation results more accurate and diverse. Extensive offline experiments on a multi-billion-scale industrial dataset and online A/B tests have demonstrated the effectiveness of TWIN-V2. Under an efficient deployment framework, TWIN-V2 has been successfully deployed to the primary traffic that serves hundreds of millions of daily active users at Kuaishou.

Life-long user behavior modeling, i.e., extracting a user's hidden interests from rich historical behaviors in months or even years, plays a central role in modern CTR prediction systems. Conventional algorithms mostly follow two cascading stages: a simple General Search Unit (GSU) for fast and coarse search over tens of thousands of long-term behaviors and an Exact Search Unit (ESU) for effective Target Attention (TA) over the small number of finalists from GSU. Although efficient, existing algorithms mostly suffer from a crucial limitation: the inconsistent target-behavior relevance metrics between GSU and ESU. As a result, their GSU usually misses highly relevant behaviors but retrieves ones considered irrelevant by ESU. In such case, the TA in ESU, no matter how attention is allocated, mostly deviates from the real user interests and thus degrades the overall CTR prediction accuracy. To address such inconsistency, we propose TWo-stage Interest Network (TWIN), where our Consistency-Preserved GSU (CP-GSU) adopts the identical target-behavior relevance metric as the TA in ESU, making the two stages twins. Specifically, to break TA's computational bottleneck and extend it from ESU to GSU, or namely from behavior length 102 to length 104 - 105, we build a novel attention mechanism by behavior feature splitting. For the video inherent features of a behavior, we calculate their linear projection by efficient pre-computing & caching strategies. And for the user-item cross features, we compress each into a one-dimentional bias term in the attention score calculation to save the computational cost. The consistency between two stages, together with the effective TA-based relevance metric in CP-GSU, contributes to significant performance gain in CTR prediction. Offline experiments on a 46 billion scale real production dataset from Kuaishou and an Online A/B test show that TWIN outperforms all compared SOTA algorithms. With optimized online infrastructure, we reduce the computational bottleneck by 99.3%, which contributes to the successful deployment of TWIN on Kuaishou, serving the main traffic of hundreds of millions of active users everyday.

With the increase of content pages and interactive buttons in online services such as online-shopping and video-watching websites, industrial-scale recommender systems face challenges in multi-domain and multi-task recommendations. The core of multi-task and multi-domain recommendation is to accurately capture user interests in multiple scenarios given multiple user behaviors. In this paper, we propose a plug-and-play Parameter and Embedding Personalized Network (PEPNet) for multi-domain and multi-task recommendation. PEPNet takes personalized prior information as input and dynamically scales the bottom-level Embedding and top-level DNN hidden units through gate mechanisms. Embedding Personalized Network (EPNet) performs personalized selection on Embedding to fuse features with different importance for different users in multiple domains. Parameter Personalized Network (PPNet) executes personalized modification on DNN parameters to balance targets with different sparsity for different users in multiple tasks. We have made a series of special engineering optimizations combining the Kuaishou training framework and the online deployment environment. By infusing personalized selection of Embedding and personalized modification of DNN parameters, PEPNet tailored to the interests of each individual obtains significant performance gains, with online improvements exceeding 1% in multiple task metrics across multiple domains. We have deployed PEPNet in Kuaishou apps, serving over 300 million users every day.

Multimodal Machine Comprehension ($\rm M^3C$) has been a challenging task that requires understanding both language and vision, as well as their integration and interaction. For example, the RecipeQA challenge, which provides several $\rm M^3C$ tasks, requires deep neural models to understand textual instructions, images of different steps, as well as the logic orders of food cooking. To address this challenge, we propose a Multi-Level Multi-Modal Transformer (MLMM-Trans) framework to integrate and understand multiple textual instructions and multiple images. Our model can conduct intensive attention mechanism at multiple levels of objects (e.g., step level and passage-image level) for sequences of different modalities. Experiments have shown that our model can achieve the state-of-the-art results on the three multimodal tasks of RecipeQA.

Multi-Choice Reading Comprehension~(MCRC) is an essential task where a machine selects the correct answer from multiple choices given a context document and a corresponding question. Existing methods usually make predictions based on a single-round reasoning process with the attention mechanism, however, this may be insufficient for tasks that require a more complex reasoning process. To effectively comprehend the context and select the correct answer from different perspectives, we propose the Read-Attend-Exclude (RAE) model which is motivated by what human readers do for MCRC in multi-rounds reasoning process. Specifically, the RAE model includes four components: the Scan Reading Module, the Attended Intensive Reading Module, the Answer Exclusion Module, and the Gated Fusion Module that makes the final decisions collectively based on the aforementioned three modules. Extensive experiments demonstrate the strong results of the proposed model on the DREAM dataset and the effectiveness of all proposed modules.

In this paper, we study the machine reading comprehension of temporal order in text. Given a document of instruction sequences, a model aims to find out the most coherent sequences of activities matching the document among all answer candidates. To tackle the task, we proposeOrdMatch model, which is able to match each activity in a sequence to the corresponding instruction in the document and regularizes the partial order of activities to match the order of instructions. We evaluate the task using the RecipeQA dataset, which includes step-by-step instructions of cooking recipes. Our model outperforms the state-of-the-art models with a wide margin. The experimental results demonstrate the effectiveness of our novel ordering regularizer. Our code will be made available at \hrefhttps://github.com/Aolius/OrdMatch https://github.com/Aolius/OrdMatch.