Kairui Fu

Kairui Fu, Tao Zhang, Shuwen Xiao et al.

Semantic identifiers (SIDs) have gained increasing attention in generative retrieval (GR) for recommendation due to their meaningful semantic discriminability. However, current studies in this field primarily (1) offer limited investigation into the construction strategies for better SIDs, and (2) their SID assessment typically relies on costly GR training. To address these challenges, we propose FORGE, a comprehensive benchmark for FOrming semantic identifieRs for Generative rEtrieval. Specifically, FORGE provides a taxonomy of the SID construction process from several perspectives and validates their impact on downstream GR through offline experiments across diverse settings. Notably, these empirical findings have led to a 0.35% increase in transaction count via online A/B experiments in the Guess You Like section of Taobao. The corresponding SID construction strategies have since been deployed at full scale on Taobao, demonstrating their practical effectiveness. To avoid expensive SID assessment that requires full GR training, we propose two novel SID evaluation metrics that are highly correlated with recommendation performance, enabling convenient evaluations without any GR training. Furthermore, to facilitate the community, we release AL-GR, the industrial dataset used in our experiments, comprising 14 billion interactions and 250 million items with the corresponding multimodal features collected from Taobao. All the code and data are available at https://github.com/selous123/al_sid.

Qihang Yu, Kairui Fu, Zhaocheng Du et al.

The scaling law, which indicates that model performance improves with increasing dataset and model capacity, has fueled a growing trend in expanding recommendation models in both industry and academia. However, the advent of large-scale recommenders also brings significantly higher computational costs, particularly under the long-sequence dependencies inherent in the user intent of recommendation systems. Current approaches often rely on pre-storing the intermediate states of the past behavior for each user, thereby reducing the quadratic re-computation cost for the following requests. Despite their effectiveness, these methods often treat memory merely as a medium for acceleration, without adequately considering the space overhead it introduces. This presents a critical challenge in real-world recommendation systems with billions of users, each of whom might initiate thousands of interactions and require massive memory for state storage. Fortunately, there have been several memory management strategies examined for compression in LLM, while most have not been evaluated on the recommendation task. To mitigate this gap, we introduce MALLOC, a comprehensive benchmark for memory-aware long sequence compression. MALLOC presents a comprehensive investigation and systematic classification of memory management techniques applicable to large sequential recommendations. These techniques are integrated into state-of-the-art recommenders, enabling a reproducible and accessible evaluation platform. Through extensive experiments across accuracy, efficiency, and complexity, we demonstrate the holistic reliability of MALLOC in advancing large-scale recommendation. Code is available at https://anonymous.4open.science/r/MALLOC.

Qihang Yu, Kairui Fu, Shengyu Zhang et al.

Recent advances in large language models (LLMs) have enabled more semantic-aware recommendations through natural language generation. Existing LLM for recommendation (LLM4Rec) methods mostly operate in a System 1-like manner, relying on superficial features to match similar items based on click history, rather than reasoning through deeper behavioral logic. This often leads to superficial and erroneous recommendations. Motivated by this, we propose ThinkRec, a thinking-based framework that shifts LLM4Rec from System 1 to System 2 (rational system). Technically, ThinkRec introduces a thinking activation mechanism that augments item metadata with keyword summarization and injects synthetic reasoning traces, guiding the model to form interpretable reasoning chains that consist of analyzing interaction histories, identifying user preferences, and making decisions based on target items. On top of this, we propose an instance-wise expert fusion mechanism to reduce the reasoning difficulty. By dynamically assigning weights to expert models based on users' latent features, ThinkRec adapts its reasoning path to individual users, thereby enhancing precision and personalization. Extensive experiments on real-world datasets demonstrate that ThinkRec significantly improves the accuracy and interpretability of recommendations. Our implementations are available in anonymous Github: https://github.com/Yu-Qi-hang/ThinkRec.

Tianyu Zhan, Kairui Fu, Chengfei Lv et al.

Generative Recommendation (GR) has recently transitioned from atomic item-indexing to Semantic ID (SID)-based frameworks to capture intrinsic item relationships and enhance generalization. However, the adoption of high-granularity SIDs leads to two critical challenges: prohibitive training overhead due to sequence expansion and unstable performance reliability characterized by non-monotonic accuracy fluctuations. We identify that these disparate issues are fundamentally rooted in the Semantic Dilution Effect, where redundant tokens waste massive computation and dilute the already sparse learning signals in recommendation. To counteract this, we propose STAMP (Semantic Trimming and Auxiliary Multi-step Prediction), a framework utilizing a dual-end optimization strategy. We argue that effective SID learning requires simultaneously addressing low input information density and sparse output supervision. On the input side, Semantic Adaptive Pruning (SAP) dynamically filters redundancy during the forward pass, converting noise-laden sequences into compact, information-rich representations. On the output side, Multi-step Auxiliary Prediction (MAP) employs a multi-token objective to densify feedback, strengthening long-range dependency capture and ensuring robust learning signals despite compressed inputs. Unifying input purification and signal amplification, STAMP enhances both training efficiency and representation capability. Experiments on public Amazon and large-scale industrial datasets show STAMP achieves 1.23--1.38$\times$ speedup and 17.2\%--54.7\% VRAM reduction while maintaining or improving performance across multiple architectures.

Tianyu Zhan, Kairui Fu, Zheqi Lv et al.

Generative recommendation systems typically leverage Semantic Identifiers (SIDs), which represent each item as a sequence of tokens that encode semantic information. However, representing item ID with multiple SIDs significantly increases input sequence length, which is a major determinant of computational complexity and memory consumption. While existing efforts primarily focus on optimizing attention computation and KV cache, we propose RASTP (Representation-Aware Semantic Token Pruning), which directly prunes less informative tokens in the input sequence. Specifically, RASTP evaluates token importance by combining semantic saliency, measured via representation magnitude, and attention centrality, derived from cumulative attention weights. Since RASTP dynamically prunes low-information or irrelevant semantic tokens, experiments on three real-world Amazon datasets show that RASTP reduces training time by 26.7\%, while maintaining or slightly improving recommendation performance. The code has been open-sourced at https://github.com/Yuzt-zju/RASTP.

Kunxi Li, Tianyu Zhan, Kairui Fu et al.

In this study, we focus on heterogeneous knowledge transfer across entirely different model architectures, tasks, and modalities. Existing knowledge transfer methods (e.g., backbone sharing, knowledge distillation) often hinge on shared elements within model structures or task-specific features/labels, limiting transfers to complex model types or tasks. To overcome these challenges, we present MergeNet, which learns to bridge the gap of parameter spaces of heterogeneous models, facilitating the direct interaction, extraction, and application of knowledge within these parameter spaces. The core mechanism of MergeNet lies in the parameter adapter, which operates by querying the source model's low-rank parameters and adeptly learning to identify and map parameters into the target model. MergeNet is learned alongside both models, allowing our framework to dynamically transfer and adapt knowledge relevant to the current stage, including the training trajectory knowledge of the source model. Extensive experiments on heterogeneous knowledge transfer demonstrate significant improvements in challenging settings, where representative approaches may falter or prove less applicable.

Kairui Fu, Zheqi Lv, Shengyu Zhang et al.

In cloud-centric recommender system, regular data exchanges between user devices and cloud could potentially elevate bandwidth demands and privacy risks. On-device recommendation emerges as a viable solution by performing reranking locally to alleviate these concerns. Existing methods primarily focus on developing local adaptive parameters, while potentially neglecting the critical role of tailor-made model architecture. Insights from broader research domains suggest that varying data distributions might favor distinct architectures for better fitting. In addition, imposing a uniform model structure across heterogeneous devices may result in risking inefficacy on less capable devices or sub-optimal performance on those with sufficient capabilities. In response to these gaps, our paper introduces Forward-OFA, a novel approach for the dynamic construction of device-specific networks (both structure and parameters). Forward-OFA employs a structure controller to selectively determine whether each block needs to be assembled for a given device. However, during the training of the structure controller, these assembled heterogeneous structures are jointly optimized, where the co-adaption among blocks might encounter gradient conflicts. To mitigate this, Forward-OFA is designed to establish a structure-guided mapping of real-time behaviors to the parameters of assembled networks. Structure-related parameters and parallel components within the mapper prevent each part from receiving heterogeneous gradients from others, thus bypassing the gradient conflicts for coupled optimization. Besides, direct mapping enables Forward-OFA to achieve adaptation through only one forward pass, allowing for swift adaptation to changing interests and eliminating the requirement for on-device backpropagation. Experiments on real-world datasets demonstrate the effectiveness and efficiency of Forward-OFA.

Tianqi Liu, Kairui Fu, Shengyu Zhang et al.

With the advancement of mobile device capabilities, deploying reranking models directly on devices has become feasible, enabling real-time contextual recommendations. When migrating models from cloud to devices, resource heterogeneity inevitably necessitates model compression. Recent quantization methods show promise for efficient deployment, yet they overlook device-specific user interests, resulting in compromised recommendation accuracy. While on-device finetuning captures personalized user preference, it imposes additional computational burden through local retraining. To address these challenges, we propose a framework for \underline{\textbf{C}}ustomizing \underline{\textbf{H}}ybrid-precision \underline{\textbf{O}}n-device model for sequential \underline{\textbf{R}}ecommendation with \underline{\textbf{D}}evice-cloud collaboration (\textbf{CHORD}), leveraging channel-wise mixed-precision quantization to simultaneously achieve personalization and resource-adaptive deployment. CHORD distributes randomly initialized models across heterogeneous devices and identifies user-specific critical parameters through auxiliary hypernetwork modules on the cloud. Our parameter sensitivity analysis operates across multiple granularities (layer, filter, and element levels), enabling precise mapping from user profiles to quantization strategy. Through on-device mixed-precision quantization, CHORD delivers dynamic model adaptation and accelerated inference without backpropagation, eliminating costly retraining cycles. We minimize communication overhead by encoding quantization strategies using only 2 bits per channel instead of 32-bit weights. Experiments on three real-world datasets with two popular backbones (SASRec and Caser) demonstrate the accuracy, efficiency, and adaptivity of CHORD.

Zheqi Lv, Wenqiao Zhang, Kairui Fu et al.

The on-device real-time data distribution shift on devices challenges the generalization of lightweight on-device models. This critical issue is often overlooked in current research, which predominantly relies on data-intensive and computationally expensive fine-tuning approaches. To tackle this, we introduce Persona, a novel personalized method using a prototype-based, backpropagation-free parameter editing framework to enhance model generalization without post-deployment retraining. Persona employs a neural adapter in the cloud to generate a parameter editing matrix based on real-time device data. This matrix adeptly adapts on-device models to the prevailing data distributions, efficiently clustering them into prototype models. The prototypes are dynamically refined via the parameter editing matrix, facilitating efficient evolution. Furthermore, the integration of cross-layer knowledge transfer ensures consistent and context-aware multi-layer parameter changes and prototype assignment. Extensive experiments on vision task and recommendation task on multiple datasets confirm Persona's effectiveness and generality.

Kairui Fu, Shengyu Zhang, Zheqi Lv et al.

Due to the continuously improving capabilities of mobile edges, recommender systems start to deploy models on edges to alleviate network congestion caused by frequent mobile requests. Several studies have leveraged the proximity of edge-side to real-time data, fine-tuning them to create edge-specific models. Despite their significant progress, these methods require substantial on-edge computational resources and frequent network transfers to keep the model up to date. The former may disrupt other processes on the edge to acquire computational resources, while the latter consumes network bandwidth, leading to a decrease in user satisfaction. In response to these challenges, we propose a customizeD slImming framework for incompatiblE neTworks(DIET). DIET deploys the same generic backbone (potentially incompatible for a specific edge) to all devices. To minimize frequent bandwidth usage and storage consumption in personalization, DIET tailors specific subnets for each edge based on its past interactions, learning to generate slimming subnets(diets) within incompatible networks for efficient transfer. It also takes the inter-layer relationships into account, empirically reducing inference time while obtaining more suitable diets. We further explore the repeated modules within networks and propose a more storage-efficient framework, DIETING, which utilizes a single layer of parameters to represent the entire network, achieving comparably excellent performance. The experiments across four state-of-the-art datasets and two widely used models demonstrate the superior accuracy in recommendation and efficiency in transmission and storage of our framework.