Alignment of Large Language Models (LLMs) aims to align outputs with human preferences, and personalized alignment further adapts models to individual users. This relies on personalized reward models that capture user-specific preferences and automatically provide individualized feedback. However, developing these models faces two critical challenges: the scarcity of feedback from individual users and the need for efficient adaptation to unseen users. We argue that addressing these constraints requires a paradigm shift from fitting static user models to ''learning to learn'' adaptation. To realize this, we propose Meta Reward Modeling (MRM), which reformulates personalized reward modeling as a meta-learning problem. Specifically, we represent each user's reward model as a weighted combination of base reward functions, and optimize the initialization of these weights using a Model-Agnostic Meta-Learning (MAML)-style framework to support fast adaptation under limited feedback. To ensure robustness, we introduce the Robust Personalization Objective (RPO), which places greater emphasis on hard-to-learn users during meta optimization. Extensive experiments on personalized preference datasets validate that MRM enhances few-shot personalization, improves user robustness, and consistently outperforms baselines. We release code at https://github.com/ModalityDance/MRM.

Large Language Model-based generative recommendation (LLMRec) has achieved notable success, but it suffers from high inference latency due to massive computational overhead and memory pressure of KV Cache. Existing KV Cache reduction methods face critical limitations: cache compression offers marginal acceleration given recommendation tasks' short decoding steps, while prompt compression risks discarding vital interaction history. Through systematic analysis of attention patterns in LLMRec, we uncover two pivotal insights: 1) layer-wise attention sparsity inversion where early layers retain dense informative patterns while later layers exhibit high redundancy, and 2) dual attention sinks phenomenon where attention scores concentrate on both head and tail tokens of input sequences. Motivated by these insights, we propose EARN, an efficient inference framework that leverages the early layers to compress information into register tokens placed at the input sequence boundaries, then focuses solely on these tokens in the subsequent layers. Extensive experiments on three datasets, two LLMRec methods and two LLM architectures demonstrate EARN's superiority, achieving up to 3.79x speedup and 80.8% KV Cache reduction with better accuracy than the general finetuning approach. Our work bridges the efficiency-effectiveness gap in LLMRec, offering practical deployment advantages for industrial scenarios.

Text-to-image retrieval is a fundamental task in multimedia retrieval.Traditional studies have typically approached this task as a discriminative problem, matching the text and image via the cross-attention mechanism (one-tower framework) or in a common embedding space (two-tower framework).The one-tower framework excels in effectiveness but falls short in efficiency, whereas the two-tower framework is efficient but struggles to maintain competitive effectiveness.In this study, we aim to enhance both effectiveness and efficiency by transforming the text-to-image retrieval task into a token-to-voken generation problem, where fine-grained interactions are incorporated to improve effectiveness while maintaining high efficiency.Despite its potential advantages, this paradigm shift presents significant challenges: 1) misalignment with high-level semantics and 2) learning gap towards the retrieval target.To address the challenges, we propose AVG, which discretizes images into vokens while aligning with both the visual information and high-level semantics.Additionally, to bridge the learning gap between generative training and the retrieval target, AVG incorporates discriminative training to modify the learning direction during token-to-voken training.Experiments demonstrate that the benefits of paradigm innovation are realized: compared with the classical two-tower method, CLIP, AVG achieves the 7.53% relative effectiveness improvement and also 4× efficiency improvement.We release code at the GitHub repository.

Generative retrieval reformulates retrieval as an autoregressive generation task, where large language models (LLMs) generate target documents directly from a query. As a novel paradigm, the mechanisms that underpin its performance and scalability remain largely unexplored. We systematically investigate training and inference scaling laws in generative retrieval, exploring how model size, training data scale, and inference-time compute jointly influence performance. We propose a novel evaluation metric inspired by contrastive entropy and generation loss, providing a continuous performance signal that enables robust comparisons across diverse generative retrieval methods. Our experiments show that n-gram-based methods align strongly with training and inference scaling laws. We find that increasing model size, training data scale, and inference-time compute all contribute to improved performance, highlighting the complementary roles of these factors in enhancing generative retrieval. Across these settings, LLaMA models consistently outperform T5 models, suggesting a particular advantage for larger decoder-only models in generative retrieval. Our findings underscore that model sizes, data availability, and inference computation interact to unlock the full potential of generative retrieval, offering new insights for designing and optimizing future systems. We release code at SLGR GitHub repository.

Web agents have emerged as a promising direction to automate Web task completion based on user instructions, significantly enhancing user experience. Recently, Web agents have evolved from traditional agents to Large Language Models (LLMs)-based Web agents. Despite their success, existing LLM-based Web agents overlook the importance of personalized data (e.g., user profiles and historical Web behaviors) in assisting the understanding of users' personalized instructions and executing customized actions.

Utilizing powerful Large Language Models (LLMs) for generative recommendation has attracted much attention. Nevertheless, a crucial challenge is transforming recommendation data into the language space of LLMs through effective item tokenization. Current approaches, such as ID, textual, and codebook-based identifiers, exhibit shortcomings in encoding semantic information, incorporating collaborative signals, or handling code assignment bias. To address these limitations, we propose LETTER (a LEarnable Tokenizer for generaTivE Recommendation), which integrates hierarchical semantics, collaborative signals, and code assignment diversity to satisfy the essential requirements of identifiers. LETTER incorporates Residual Quantized VAE for semantic regularization, a contrastive alignment loss for collaborative regularization, and a diversity loss to mitigate code assignment bias. We instantiate LETTER on two models and propose a ranking-guided generation loss to augment their ranking ability theoretically. Experiments on three datasets validate the superiority of LETTER, advancing the state-of-the-art in the field of LLM-based generative recommendation.

Harnessing Large Language Models (LLMs) for recommendation is rapidly emerging, which relies on two fundamental steps to bridge the recommendation item space and the language space: 1) item indexing utilizes identifiers to represent items in the language space, and 2) generation grounding associates LLMs' generated token sequences to in-corpus items. However, previous methods exhibit inherent limitations in the two steps. Existing ID-based identifiers (e.g., numeric IDs) and description-based identifiers (e.g., titles) either lose semantics or lack adequate distinctiveness. Moreover, prior generation grounding methods might generate invalid identifiers, thus misaligning with in-corpus items. To address these issues, we propose a novel Transition paradigm for LLM-based Recommender (named TransRec) to bridge items and language. Specifically, TransRec presents multi-facet identifiers, which simultaneously incorporate ID, title, and attribute for item indexing to pursue both distinctiveness and semantics. Additionally, we introduce a specialized data structure for TransRec to ensure generating valid identifiers only and utilize substring indexing to encourage LLMs to generate from any position of identifiers. Lastly, TransRec presents an aggregated grounding module to leverage generated multi-facet identifiers to rank in-corpus items efficiently. We instantiate TransRec on two backbone models, BART-large and LLaMA-7B.

Leveraging Large Language Models (LLMs) for recommendation has recently garnered considerable attention, where fine-tuning plays a key role in LLMs' adaptation. However, the cost of fine-tuning LLMs on rapidly expanding recommendation data limits their practical application. To address this challenge, few-shot fine-tuning offers a promising approach to quickly adapt LLMs to new recommendation data. We propose the task of data pruning for efficient LLM-based recommendation, aimed at identifying representative samples tailored for LLMs' few-shot fine-tuning. While coreset selection is closely related to the proposed task, existing coreset selection methods often rely on suboptimal heuristic metrics or entail costly optimization on large-scale recommendation data.

In recent years, conversational agents have provided a natural and convenient access to useful information in people’s daily life, along with a broad and new research topic, conversational question answering (QA). On the shoulders of conversational QA, we study the conversational open-domain QA problem, where users’ information needs are presented in a conversation and exact answers are required to extract from the Web. Despite its significance and value, building an effective conversational open-domain QA system is non-trivial due to the following challenges: (1) precisely understand conversational questions based on the conversation context; (2) extract exact answers by capturing the answer dependency and transition flow in a conversation; and (3) deeply integrate question understanding and answer extraction. To address the aforementioned issues, we propose an end-to-end Dynamic Graph Reasoning approach to Conversational open-domain QA (DGRCoQA for short). DGRCoQA comprises three components, i.e., a dynamic question interpreter (DQI), a graph reasoning enhanced retriever (GRR), and a typical Reader, where the first one is developed to understand and formulate conversational questions while the other two are responsible to extract an exact answer from the Web. In particular, DQI understands conversational questions by utilizing the QA context, sourcing from predicted answers returned by the Reader, to dynamically attend to the most relevant information in the conversation context. Afterwards, GRR attempts to capture the answer flow and select the most possible passage that contains the answer by reasoning answer paths over a dynamically constructed context graph . Finally, the Reader, a reading comprehension model, predicts a text span from the selected passage as the answer. DGRCoQA demonstrates its strength in the extensive experiments conducted on a benchmark dataset. It significantly outperforms the existing methods and achieves the state-of-the-art performance.

Recent years have witnessed a flourishing of community-driven question answering (cQA), like Yahoo! Answers and AnswerBag, where people can seek precise information. After 2010, some novel cQA systems, including Quora and Zhihu, gained momentum. Besides interactions, the latter enables users to label the questions with topic tags that highlight the key points conveyed in the questions. In this article, we shed light on automatically annotating a newly posted question with topic tags that are predefined and preorganized into a directed acyclic graph. To accomplish this task, we present an end-to-end deep interactive embedding model to jointly learn the embeddings of questions and topics by projecting them into the same space for a similarity measure. In particular, we first learn the embeddings of questions and topic tags by two deep parallel models. Thereinto, we regularize the embeddings of topic tags via fully exploring their hierarchical structures, which is able to alleviate the problem of imbalanced topic distribution. Thereafter, we interact each question embedding with the topic tag matrix, i.e., all the topic tag embeddings. Following that, a sigmoid cross-entropy loss is appended to reward the positive question-topic pairs and penalize the negative ones. To justify our model, we have conducted extensive experiments on an unprecedented large-scale social QA dataset obtained from Zhihu.com, and the experimental results demonstrate that our model achieves superior performance to several state-of-the-art baselines.