Deep-research agents are capable of executing multi-step web exploration, targeted retrieval, and sophisticated question answering. Despite their powerful capabilities, deep-research agents face two critical bottlenecks: (1) the lack of large-scale, challenging datasets with real-world difficulty, and (2) the absence of accessible, open-source frameworks for data synthesis and agent training. To bridge these gaps, we first construct DeepResearch-9K, a large-scale challenging dataset specifically designed for deep-research scenarios built from open-source multi-hop question-answering (QA) datasets via a low-cost autonomous pipeline. Notably, it consists of (1) 9000 questions spanning three difficulty levels from L1 to L3 (2) high-quality search trajectories with reasoning chains from Tongyi-DeepResearch-30B-A3B, a state-of-the-art deep-research agent, and (3) verifiable answers. Furthermore, we develop an open-source training framework DeepResearch-R1 that supports (1) multi-turn web interactions, (2) different reinforcement learning (RL) approaches, and (3) different reward models such as rule-based outcome reward and LLM-as-Judge feedback. Finally, empirical results demonstrate that agents trained on DeepResearch-9K under our DeepResearch-R1 achieve state-of-the-art results on challenging deep-research benchmarks. We release the DeepResearch-9K dataset on https://huggingface.co/datasets/artillerywu/DeepResearch-9K and the code of DeepResearch-R1 on https://github.com/Applied-Machine-Learning-Lab/SIGIR2026_DeepResearch-R1.

Retrieval-Augmented Generation (RAG) utilizes external knowledge to augment Large Language Models’ (LLMs) reliability. For flexibility, agentic RAG employs autonomous, multi-round retrieval and reasoning to resolve queries. Although recent agentic RAG has improved via reinforcement learning, they often incur substantial token overhead from search and reasoning. This tradeoff prioritizes accuracy over efficiency. To address this issue, this work proposes TeaRAG, a T oken- e fficient a gentic RAG framework capable of compressing both retrieval content and reasoning steps. (1) First, the retrieved content is compressed by augmenting chunk-based semantic retrieval with a graph retrieval using concise triplets. A knowledge association graph is then built from semantic similarity and co-occurrence. Finally, Personalized PageRank is leveraged to highlight key knowledge within this graph, reducing the number of tokens per retrieval. (2) Besides, to reduce reasoning steps, Iterative Process-aware Direct Preference Optimization (IP-DPO) is proposed. Specifically, our reward function evaluates the knowledge sufficiency by a knowledge matching mechanism, while penalizing excessive reasoning steps. This design can produce high-quality preference-pair datasets, supporting iterative DPO to improve reasoning conciseness. Across six datasets, TeaRAG improves the average Exact Match by \(4\%\) and \(2\%\) while reducing output tokens by \(61\%\) and \(59\%\) on Llama3-8B-Instruct and Qwen2.5-14B-Instruct, respectively. Code is available at https://github.com/Applied-Machine-Learning-Lab/TeaRAG .

Sequential recommendation (SR) aims to capture users' dynamic interests and sequential patterns based on their historical interactions. Recently, the powerful capabilities of large language models (LLMs) have driven their adoption in SR. However, we identify two critical challenges in existing LLM-based SR methods: 1) embedding collapse when incorporating pre-trained collaborative embeddings and 2) catastrophic forgetting of quantized embeddings when utilizing semantic IDs. These issues dampen the model scalability and lead to suboptimal recommendation performance. Therefore, based on LLMs like Llama3-8B-instruct, we introduce a novel SR framework named MME-SID, which integrates multimodal embeddings and quantized embeddings to mitigate embedding collapse. Additionally, we propose a Multimodal Residual Quantized Variational Autoencoder (MM-RQ-VAE) with maximum mean discrepancy as the reconstruction loss and contrastive learning for alignment, which effectively preserve intra-modal distance information and capture inter-modal correlations, respectively. To further alleviate catastrophic forgetting, we initialize the model with the trained multimodal code embeddings. Finally, we fine-tune the LLM efficiently using LoRA in a multimodal frequency-aware fusion manner. Extensive experiments on three public datasets validate the superior performance of MME-SID thanks to its capability to mitigate embedding collapse and catastrophic forgetting. The implementation code and datasets are publicly available for reproduction: https://github.com/Applied-Machine-Learning-Lab/MME-SID.

Multi-Scenario Recommendation (MSR) tasks, referring to building a unified model to enhance performance across all recommendation scenarios, have recently gained considerable attention. However, current research in MSR faces two significant challenges that hinder the field's development: the absence of uniform procedures for multi-scenario dataset processing, thus hindering fair comparisons, and most models being closed-source, which complicates comparisons with current SOTA models. Consequently, we introduce our benchmark, Scenario-Wise Rec, which comprises six public datasets and twelve baseline models, along with a training and evaluation pipeline. We further validate Scenario-Wise Rec on an industrial advertising dataset, underscoring its robustness. We hope the benchmark will give researchers clear insights into prior work, enabling them to develop novel models and thereby fostering a collaborative research ecosystem in MSR. Our source code is publicly available (https://github.com/Applied-Machine-Learning-Lab/Scenario-Wise-Rec).

Feature selection is crucial in recommender systems for improving model efficiency and predictive performance. Conventional approaches typically employ surrogate models-such as decision trees or neural networks-to estimate feature importance. However, their effectiveness is inherently constrained, as these models may struggle under suboptimal training conditions, including feature collinearity, high-dimensional sparsity, and insufficient data. In this paper, we propose SELF, a SurrogatE-Light Feature selection method for deep recommender systems. SELF integrates semantic reasoning from Large Language Models (LLMs) with task-specific learning from surrogate models, enabling an automated and lightweight feature selection process. Specifically, LLMs first produce a semantically informed ranking of feature importance, which is subsequently refined by a surrogate model, effectively integrating general world knowledge with task-specific learning. Comprehensive experiments on three public datasets from real-world recommender platforms validate the effectiveness of SELF. To facilitate reproducibility, our code is publicly available.

Large language models (LLMs) have demonstrated their capabilities across various natural language processing (NLP) tasks. Their potential in e-commerce is also substantial, evidenced by existing implementations in scenarios such as platform search and recommender systems. One obstinate concern associated with LLMs is the factuality issue (e.g., hallucination), which is urgent in e-commerce due to its significant impact on user experience and revenue. While some methods aim to evaluate the factuality of LLMs, issues such as lack of objectivity, high consumption, and lack of domain expertise arise. To this end, leveraging a collected knowledge graph (KG) as a reliable source, we propose ECKGBench, a question-answering dataset to assess LLMs' capacity in e-commerce. Specifically, each question is automatically generated based on one KG triple through a standardized pipeline, guaranteeing evaluation quality and reliability. We evaluate advanced LLMs using ECKGBench and provide insights into experimental results. The dataset is available online at~ https://github.com/OpenStellarTeam/ECKGBench.

Sequential recommendation aims to capture user preferences by modeling sequential patterns in user-item interactions. However, these models are often influenced by noise such as accidental interactions, leading to suboptimal performance. Therefore, to reduce the effect of noise, some works propose explicitly identifying and removing noisy items. However, we find that simply relying on collaborative information may result in an over-denoising problem, especially for cold items. To overcome these limitations, we propose a novel framework: Interest Alignment for Denoising Sequential Recommendation (IADSR) which integrates both collaborative and semantic information. Specifically, IADSR is comprised of two stages: in the first stage, we obtain the collaborative and semantic embeddings of each item from a traditional sequential recommendation model and an LLM, respectively. In the second stage, we align the collaborative and semantic embeddings and then identify noise in the interaction sequence based on long-term and short-term interests captured in the collaborative and semantic modalities. Our extensive experiments on four public datasets validate the effectiveness of the proposed framework and its compatibility with different sequential recommendation systems. The code and data are released for reproducibility: https://github.com/Applied-Machine-Learning-Lab/IADSR.

Due to exceptional reasoning and understanding abilities, the Large Language Model (LLM) has revolutionized the pattern of many fields, including recommender systems (RS). There has been a handful of research that focuses on empowering the RS by LLM. Recently, considering the latency and memory costs in real-world applications, LLM-Enhanced RS (LLMERS) is highlighted. This direction pushes the LLM into the online system with a large step by eliminating the utilization of LLM during inference. As a cutting-edge field, there is a clear need for a comprehensive survey to summarize this direction. In this survey, we systematically investigate the most up-to-date works of LLM-enhanced RS to boost this direction. Based on the component of an RS model that the LLM aims to augment, the basic taxonomy includes Knowledge Enhancement, Interaction Enhancement and Model Enhancement. Additionally, we identify several promising research directions. To facilitate access to the surveyed papers, we release a repository.

Sequential Recommendation (SR) aims to leverage the sequential patterns in users' historical interactions to accurately track their preferences. However, the primary reliance of existing SR methods on collaborative data results in challenges such as the cold-start problem and sub-optimal performance. Concurrently, despite the proven effectiveness of large language models (LLMs), their integration into commercial recommender systems is impeded by issues such as high inference latency, incomplete capture of all distribution statistics, and catastrophic forgetting. To address these issues, we introduce a novel Pre-train, Align, and Disentangle (PAD) framework to enhance SR models with LLMs. In particular, we initially pre-train both the SR and LLM models to obtain collaborative and textual embeddings. Subsequently, we propose a characteristic recommendation-anchored alignment loss using multi-kernel maximum mean discrepancy with Gaussian kernels. Lastly, a triple-experts architecture, comprising aligned and modality-specific experts with disentangled embeddings, is fine-tuned in a frequency-aware manner. Experimental results on three public datasets validate the efficacy of PAD, indicating substantial enhancements and compatibility with various SR backbone models, particularly for cold items. The code and datasets are accessible for reproduction: https://github.com/Applied-Machine-Learning-Lab/PAD.

As the demand for more personalized recommendation grows and a dramatic boom in commercial scenarios arises, the study on multi-scenario recommendation (MSR) has attracted much attention, which uses the data from all scenarios to simultaneously improve their recommendation performance. However, existing methods tend to integrate insufficient scenario knowledge and neglect learning personalized cross-scenario preferences, thus leading to sub-optimal performance. Meanwhile, though large language model (LLM) has shown great capability of reasoning and capturing semantic information, the high inference latency and high computation cost of tuning hinder its implementation in industrial recommender systems. To fill these gaps, we propose an LLM-enhanced paradigm LLM4MSR in this work. Specifically, we first leverage LLM to uncover multi-level knowledge from the designed scenario- and user-level prompt without fine-tuning the LLM, then adopt hierarchical meta networks to generate multi-level meta layers to explicitly improve the scenario-aware and personalized recommendation capability. Our experiments on KuaiSAR-small, KuaiSAR, and Amazon datasets validate significant advantages of LLM4MSR: (i) the effectiveness and compatibility with different multi-scenario backbone models, (ii) high efficiency and deployability on industrial recommender systems, and (iii) improved interpretability. The implemented code and data is available to ease reproduction.

In today's digital landscape, Deep Recommender Systems (DRS) play a crucial role in navigating and customizing online content for individual preferences. However, conventional methods, which mainly depend on single recommendation task, scenario, data modality and user behavior, are increasingly seen as insufficient due to their inability to accurately reflect users' complex and changing preferences. This gap underscores the need for multi-granularity modeling, which are central to overcoming these limitations by integrating diverse tasks, scenarios, modalities, and behaviors in the recommendation process, thus promising significant enhancements in recommendation precision, efficiency, and customization. In this paper, from the multi-scenario perspective, we illustrate our existing explorations and present results. Ultimately, we wish to highlight our multi-granularity approach sheds light on building the next generation of recommender system1 .

Click-Through Rate (CTR) prediction is a fundamental technique in recommendation and advertising systems. Recent studies have shown that implementing multi-scenario recommendations contributes to strengthening information sharing and improving overall performance. However, existing multi-scenario models only consider coarse-grained explicit scenario modeling that depends on pre-defined scenario identification from manual prior rules, which is biased and sub-optimal. To address these limitations, we propose a Scenario-Aware Hierarchical Dynamic Network for Multi-Scenario Recommendations (HierRec), which perceives implicit patterns adaptively, and conducts explicit and implicit scenario modeling jointly. In particular, HierRec designs a basic scenario-oriented module based on the dynamic weight to capture scenario-specific representations. Then the hierarchical explicit and implicit scenario-aware modules are proposed to model hybrid-grained scenario information, where the multi-head implicit modeling design contributes to perceiving distinctive patterns from different perspectives. Our experiments on two public datasets and real-world industrial applications on a mainstream online advertising platform demonstrate that HierRec outperforms existing models significantly. The implementation code is available for reproducibility.

With the explosive growth of various commercial scenarios, there is an increasing number of studies on multi-scenario recommendation (MSR) which trains the recommender system with the data from multiple scenarios, aiming to improve the recommendation performance on all these scenarios synchronously. However, due to the large discrepancy in the number of interactions among domains, multi-scenario recommendation models usually suffer from insufficient learning and negative transfer especially on the cold-start scenarios, thus exacerbating the data sparsity issue. To fill this gap, in this work we propose a novel diffusion model enhanced paradigm tailored for the cold-start problem in multi-scenario recommendation in a data-driven generative manner. Specifically, based on all-domain data, we leverage the diffusion model with our newly designed variance schedule and the proposed classifier, which explicitly boosts the recommendation performance on the cold-start scenarios by exploiting the generated high-quality and informative embedding, leveraging the abundance of rich scenarios. Our experiments on Douban and Amazon datasets demonstrate two strengths of the proposed paradigm: (i) its effectiveness with a significant increase of 8.5% and 1% in accuracy on the two datasets, and (ii) its compatibility with various multi-scenario backbone models. The implementation code is available for easy reproduction.

Graph Contrastive Learning (GCL) stands as a potent framework for unsupervised graph representation learning that has gained traction across numerous graph learning applications. The effectiveness of GCL relies on generating high-quality contrasting samples, enhancing the model’s ability to discern graph semantics. However, the prevailing GCL methods face two key challenges: 1) introducing noise during graph augmentations and 2) requiring additional storage for generated samples, which degrade the model performance. In this paper, we propose novel approaches, GKCL (i.e., Graph Knowledge Contrastive Learning) and DGKCL (i.e., Distilled Graph Knowledge Contrastive Learning), that leverage multi-level graph knowledge to create noise-free contrasting pairs. This framework not only addresses the noise-related challenges but also circumvents excessive storage demands. Furthermore, our method incorporates a knowledge distillation component to optimize the trained embedding tables, reducing the model’s scale while ensuring superior performance, particularly for the scenarios with smaller embedding sizes. Comprehensive experimental evaluations on three public benchmark datasets underscore the merits of our proposed method and elucidate its properties, which primarily reflect the performance of the proposed method equipped with different embedding sizes and how the distillation weight affects the overall performance.

Deep recommender systems typically involve numerous feature fields for users and items, with a large number of low-frequency features. These low-frequency features would reduce the prediction accuracy with large storage space due to their vast quantity and inadequate training. Some pioneering studies have explored embedding compression techniques to address this issue of the trade-off between storage space and model predictability. However, these methods have difficulty compacting the embedding of low-frequency features in various feature fields due to the high demand for human experience and computing resources during hyper-parameter searching. In this paper, we propose the AutoDPQ framework, which automatically compacts low-frequency feature embeddings for each feature field to an adaptive magnitude. Experimental results indicate that AutoDPQ can significantly reduce the parameter space while improving recommendation accuracy. Moreover, AutoDPQ is compatible with various deep CTR models by improving their performance significantly with high efficiency.

With the explosive growth of commercial applications of recommender systems, multi-scenario recommendation (MSR) has attracted considerable attention, which utilizes data from multiple domains to improve their recommendation performance simultaneously. However, training a unified deep recommender system (DRS) may not explicitly comprehend the commonality and difference among domains, whereas training an individual model for each domain neglects the global information and incurs high computation costs. Likewise, fine-tuning on each domain is inefficient, and recent advances that apply the prompt tuning technique to improve fine-tuning efficiency rely solely on large-sized transformers. In this work, we propose a novel prompt-enhanced paradigm for multi-scenario recommendation. Specifically, a unified DRS backbone model is first pre-trained using data from all the domains in order to capture the commonality across domains. Then, we conduct prompt tuning with two novel prompt modules, capturing the distinctions among various domains and users. Our experiments on Douban, Amazon, and Ali-CCP datasets demonstrate the effectiveness of the proposed paradigm with two noticeable strengths: (i) its great compatibility with various DRS backbone models, and (ii) its high computation and storage efficiency with only 6% trainable parameters in prompt tuning phase. The implementation code is available for easy reproduction.

In recommender systems, reinforcement learning solutions have effectively boosted recommendation performance because of their ability to capture long-term user-system interaction. However, the action space of the recommendation policy is a list of items, which could be extremely large with a dynamic candidate item pool. To overcome this challenge, we propose a hyper-actor and critic learning framework where the policy decomposes the item list generation process into a hyper-action inference step and an effect-action selection step. The first step maps the given state space into a vectorized hyper-action space, and the second step selects the item list based on the hyper-action. In order to regulate the discrepancy between the two action spaces, we design an alignment module along with a kernel mapping function for items to ensure inference accuracy and include a supervision module to stabilize the learning process. We build simulated environments on public datasets and empirically show that our framework is superior in recommendation compared to standard RL baselines.