User retention is critical in large-scale recommender systems, significantly influencing online platforms' long-term success. Existing methods typically focus on short-term engagement, neglecting the evolving dynamics of user behaviors over time. Reinforcement learning (RL) methods, though promising for optimizing long-term rewards, face challenges like delayed credit assignment and sample inefficiency. We introduce Stratified Expert Cloning (SEC), an imitation learning framework that leverages abundant interaction data from high-retention users to learn robust policies. SEC incorporates: 1) multi-level expert stratification to model diverse retention behaviors; 2) adaptive expert selection to dynamically match users with appropriate policies based on their state and retention history; and 3) action entropy regularization to enhance recommendation diversity and policy generalization. Extensive offline evaluations and online A/B tests on major video platforms (Kuaishou and Kuaishou Lite) with hundreds of millions of users validate SEC's effectiveness. Results show substantial improvements, achieving cumulative lifts of 0.098% and 0.122% in active days on the two platforms respectively, each translating into over 200,000 additional daily active users.

Visual instruction tuning is crucial for enhancing the zero-shot generalization capability of Multi-modal Large Language Models (MLLMs). In this paper, we aim to investigate a fundamental question: “what makes for good visual instructions”. Through a comprehensive empirical study, we find that instructions focusing on complex visual reasoning tasks are particularly effective in improving the performance of MLLMs, with results correlating to instruction complexity. Based on this insight, we develop a systematic approach to automatically create high-quality complex visual reasoning instructions. Our approach employs a synthesize-complicate-reformulate paradigm, leveraging multiple stages to gradually increase the complexity of the instructions while guaranteeing quality. Based on this approach, we create the ComVint dataset with 32K examples, and fine-tune four MLLMs on it. Experimental results consistently demonstrate the enhanced performance of all compared MLLMs, such as a 27.86% and 27.60% improvement for LLaVA on MME-Perception and MME-Cognition, respectively. Our code and data are publicly available at the link: https://github.com/RUCAIBox/ComVint.

Ultra-high Spatial Resolution (UHSR) Land Cover Classification is increasingly important for urban analysis, enabling fine-scale planning, ecological monitoring, and infrastructure management. It identifies land cover types on sub-meter remote sensing imagery, capturing details such as building outlines, road networks, and distinct boundaries. However, most existing methods focus on 1 m imagery and rely heavily on large-scale annotations, while UHSR data remain scarce and difficult to annotate, limiting practical applicability. To address these challenges, we introduce Baltimore Atlas, a UHSR land cover classification framework that reduces reliance on large-scale training data and delivers high-accuracy results. Baltimore Atlas builds on three key ideas: (1) Baltimore Atlas Dataset, a 0.3 m resolution dataset based on aerial imagery of Baltimore City; (2) FreqWeaver Adapter, a parameter-efficient adapter that transfers SAM2 to this domain, leveraging foundation model knowledge to reduce training data needs while enabling fine-grained detail and structural modeling; (3) Uncertainty-Aware Teacher Student Framework, a semi-supervised framework that exploits unlabeled data to further reduce training dependence and improve generalization across diverse scenes. Using only 5.96% of total model parameters, our approach achieves a 1.78% IoU improvement over existing parameter-efficient tuning strategies and a 3.44% IoU gain compared to state-of-the-art high-resolution remote sensing segmentation methods on the Baltimore Atlas Dataset.

As an effective tool for eliciting the power of Large Language Models (LLMs), prompting has recently demonstrated unprecedented abilities across a variety of complex tasks. To further improve the performance, prompt ensemble has attracted substantial interest for tackling the hallucination and instability of LLMs. However, existing methods usually adopt a two-stage paradigm, which requires a pre-prepared set of prompts with substantial manual effort, and is unable to perform directed optimization for different weak learners. In this paper, we propose a simple, universal, and automatic method named PREFER (Prompt Ensemble learning via Feedback-Reflect-Refine) to address the stated limitations. Specifically, given the fact that weak learners are supposed to focus on hard examples during boosting, PREFER builds a feedback mechanism for reflecting on the inadequacies of existing weak learners. Based on this, the LLM is required to automatically synthesize new prompts for iterative refinement. Moreover, to enhance stability of the prompt effect evaluation, we propose a novel prompt bagging method involving forward and backward thinking, which is superior to majority voting and is beneficial for both feedback and weight calculation in boosting. Extensive experiments demonstrate that our PREFER achieves state-of-the-art performance in multiple types of tasks by a significant margin. We have made our code publicly available.

Within online platforms, it is critical to capture the dynamic user preference from the sequential interaction behaviors for making accurate recommendation over time. Recently, significant progress has been made in sequential recommendation with deep learning. However, existing neural sequential recommender often suffer from the data sparsity issue in real-world applications.

Nowadays, news recommendation has become a popular channel for users to access news of their interests. How to represent rich textual contents of news and precisely match users' interests and candidate news lies in the core of news recommendation. However, existing recommendation methods merely learn textual representations from in-domain news data, which limits their generalization ability to new news that are common in cold-start scenarios. Meanwhile, many of these methods represent each user by aggregating the historically browsed news into a single vector and then compute the matching score with the candidate news vector, which may lose the low-level matching signals. In this paper, we explore the use of the successful BERT pre-training technique in NLP for news recommendation and propose a BERT-based user-news matching model, called UNBERT. In contrast to existing research, our UNBERT model not only leverages the pre-trained model with rich language knowledge to enhance textual representation, but also captures multi-grained user-news matching signals at both word-level and news-level. Extensive experiments on the Microsoft News Dataset (MIND) demonstrate that our approach constantly outperforms the state-of-the-art methods.

Personalized news recommendation is a critical technology to help users find interested news, and how to precisely match users' interests and candidate news lies in the core of news recommendation. Existing studies generally learn user's interest vector by aggregating his/her browsed news and then match it with the candidate news vector, which may lose the textual semantic matching signals for recommendation. In this paper, we propose an Attentive Multi-field Matching (AMM) framework for news recommendation which captures the semantic matching representations between each browsed news and candidate news, and then aggregates them as final user-news matching signal. In addition, our method incorporates multi-field information and designs a within-field and cross-field matching mechanism, which leverages complementary information from different fields (e.g., titles, abstracts and bodies) and obtain the multi-field matching representations. To achieve a comprehensive semantic understanding, we employ the most popular language model BERT to learn the matching representation of each browsed-candidate news pair, and incorporate the attention mechanism in aggregating procedure to characterize the importance of each matching representation for the final user-news matching signal. Experiments on the real world datasets validate the effectiveness of AMM.