Realignment becomes necessary when a language model (LM) fails to meet expected performance. We propose a flexible realignment framework that supports quantitative control of alignment degree during training and inference. This framework incorporates **Training-time Realignment (TrRa)**, which efficiently realigns the reference model by leveraging the controllable fusion of logits from both the reference and already aligned models. For example, TrRa reduces token usage by **54.63%** on DeepSeek-R1-Distill-Qwen-1.5B without any performance degradation, outperforming DeepScaleR-1.5B’s **33.86%**. To complement TrRa during inference, we introduce a layer adapter that enables **smooth Inference-time Realignment (InRa)**. This adapter is initialized to perform an identity transformation at the bottom layer and is inserted preceding the original layers. During inference, input embeddings are simultaneously processed by the adapter and the original layer, followed by the remaining layers, and then controllably interpolated at the logit level. We upgraded DeepSeek-R1-Distill-Qwen-7B from a slow-thinking model to one that supports both fast and slow thinking, allowing flexible alignment control even **during inference**. By encouraging deeper reasoning, it even surpassed its original performance.

An effective online recommendation system should jointly capture users' long-term and short-term preferences in both users' internal behaviors (from the target recommendation task) and external behaviors (from other tasks). However, it is extremely challenging to conduct fast adaptations to real-time new trends while making full use of all historical behaviors in large-scale systems, due to the real-world limitations in real-time training efficiency and external behavior acquisition. To address these practical challenges, we propose a novel Long Short-Term Temporal Meta-learning framework (LSTTM) for online recommendation. It arranges user multi-source behaviors in a global long-term graph and an internal short-term graph, and conducts different GAT-based aggregators and training strategies to learn user short-term and long-term preferences separately. To timely capture users' real-time interests, we propose a temporal meta-learning method based on MAML under an asynchronous optimization strategy for fast adaptation, which regards recommendations at different time periods as different tasks. In experiments, LSTTM achieves significant improvements on both offline and online evaluations. It has been deployed on a widely-used online recommendation system named WeChat Top Stories, affecting millions of users.

Personalized recommendation often relies on user historical behaviors to provide items for users. It is intuitive that future information also contains essential messages as supplements to user historical behaviors. However, we cannot directly encode future information into models, since we are unable to get future information in online serving. In this work, we propose a novel adversarial future encoding (AFE) framework to make full use of informative future features in different types of recommendation models. Specifically, AFE contains a future-aware discriminator and a generator. The future-aware discriminator takes both common features and future features as inputs, working as a recommendation prophet to judge user-item pairs. In contrast, the generator is considered as a challenger, which generates items with only common features, aiming to confuse the future-aware prophet. The future-aware discriminator can inspire the generator (to be deployed online) to produce better results. We further conduct a multi-factor optimization to enable a fast and stable model convergence via the direct learning and knowledge distillation losses. Moreover, we have adopted AFE on both a list-wise RL-based ranking model and a point-wise ranking model to verify its universality. In experiments, we conduct sufficient evaluations on two large-scale datasets, achieving significant improvements on both offline and online evaluations. Currently, we have deployed AFE on a real-world system, affecting millions of users. The source code is in https://github.com/modriczhang/AFE.

Integrated recommendation aims to jointly recommend heterogeneous items in the main feed from different sources via multiple channels, which needs to capture user preferences on both item and channel levels. It has been widely used in practical systems by billions of users, while few works concentrate on the integrated recommendation systematically. In this work, we propose a novel Hierarchical reinforcement learning framework for integrated recommendation (HRL-Rec), which divides the integrated recommendation into two tasks to recommend channels and items sequentially. The low-level agent is a channel selector, which generates a personalized channel list. The high-level agent is an item recommender, which recommends specific items from heterogeneous channels under the channel constraints. We design various rewards for both recommendation accuracy and diversity, and propose four losses for fast and stable model convergence. We also conduct an online exploration for sufficient training. In experiments, we conduct extensive offline and online experiments on a billion-level real-world dataset to show the effectiveness of HRL-Rec. HRL-Rec has also been deployed on WeChat Top Stories, affecting millions of users. The source codes are released in https://github.com/modriczhang/HRL-Rec.

Reinforcement learning (RL) has been verified in real-world list-wise recommendation. However, RL-based recommendation suffers from huge memory and computation costs due to its large-scale models. Knowledge distillation (KD) is an effective approach for model compression widely used in practice. However, RL-based models strongly rely on sufficient explorations on the enormous user-item space due to the data sparsity issue, which multiplies the challenges of KD with RL models. What the teacher should teach and how much the student should learn from each lesson need to be carefully designed. In this work, we propose a novel Distilled reinforcement learning framework for recommendation (DRL-Rec), which aims to improve both effectiveness and efficiency in list-wise recommendation. Specifically, we propose an Exploring and filtering module before the distillation, which decides what lessons the teacher should teach from both teachers' and students' aspects. We also conduct a Confidence-guided distillation at both output and intermediate levels with a list-wise KL divergence loss and a Hint loss, which aims to understand how much the student should learn for each lesson. We achieve significant improvements on both offline and online evaluations in a well-known recommendation system. DRL-Rec has been deployed on WeChat Top Stories for more than six months, affecting millions of users. The source codes are released in https://github.com/modriczhang/DRL-Rec.

Learning disentangled representations for user intentions from multi-feedback (i.e., positive and negative feedback) can enhance the accuracy and explainability of recommendation algorithms. However, learning such disentangled representations from multi-feedback data is challenging because i) multi-feedback is complex: there exist complex relations among different types of feedback (e.g., click, unclick, and dislike, etc) as well as various user intentions, and ii) multi-feedback is noisy: there exists noisy (useless) information both in features and labels, which may deteriorate the recommendation performance. Existing works on disentangled representation learning only focus on positive feedback, failing to handle the complex relations and noise hidden in multi-feedback data. To solve this problem, in this work we propose a Curriculum Disentangled Recommendation (CDR) model that is capable of efficiently learning disentangled representations from complex and noisy multi-feedback for better recommendation. Concretely, we design a co-filtering dynamic routing mechanism that simultaneously captures the complex relations among different behavioral feedback and user intentions as well as denoise the representations in the feature level. We then present an adjustable self-evaluating curriculum that is able to evaluate sample difficulties for better model training and conduct denoising in the label level via disregarding useless information. Our extensive experiments on several real-world datasets demonstrate that the proposed CDR model can significantly outperform several state-of-the-art methods in terms of recommendation accuracy.

Users of recommendation systems usually focus on one topic at a time. When finishing reading an item, users may want to access more relevant items related to the last read one as extended reading. However, conventional recommendation systems are hard to provide the continuous extended reading function of these relevant items, since the main recommendation results should be diversified. In this paper, we propose a new task named recommendation suggestion, which aims to (1) predict whether users want extended reading, and (2) provide appropriate relevant items as suggestions. These recommended relevant items are arranged in a relevant box and instantly inserted below the clicked item in the main feed. The challenge of recommendation suggestion on relevant items is that it should further consider semantic relevance and information gain besides CTR-related factors. Moreover, the real-time relevant box insertion may also harm the overall performance when users do not want extended reading. To address these issues, we propose a novel Real-time relevant recommendation suggestion (R3S) framework, which consists of an Item recommender and a Box trigger. We extract features from multiple aspects including feature interaction, semantic similarity and information gain as different experts, and propose a new Multi-critic multi-gate mixture-of-experts (M3oE) strategy to jointly consider different experts with multi-head critics. In experiments, we conduct both offline and online evaluations on a real-world recommendation system with detailed ablation tests. The significant improvements in item/box related metrics verify the effectiveness of R3S. Moreover, we have deployed R3S on WeChat Top Stories, which affects millions of users. The source codes are in https://github.com/modriczhang/R3S.

Real-world recommendation systems usually have different learning objectives and evaluation criteria on accuracy, diversity or novelty. Therefore, multi-objective recommendation (MOR) has been widely explored to jointly model different objectives. Pareto efficiency, where no objective can be further improved without hurting others, is viewed as an optimal situation in multi-objective optimization. Recently, Pareto efficiency model has been introduced to MOR, while all existing scalarization methods only have shared objective weights for all instances. To capture users’ objective-level preferences and enhance personalization in Pareto-efficient recommendation, we propose a novel Personalized Approximate Pareto-Efficient Recommendation (PAPERec) framework for multi-objective recommendation. Specifically, we design an approximate Pareto-efficient learning based on scalarization with KKT conditions that closely mimics Pareto efficiency, where users have personalized weights on different objectives. We propose a Pareto-oriented reinforcement learning module to find appropriate personalized objective weights for each user, with the weighted sum of multiple objectives’ gradients considered in reward. In experiments, we conduct extensive offline and online evaluations on a real-world recommendation system. The significant improvements verify the effectiveness of PAPERec in practice. We have deployed PAPERec on WeChat Top Stories, affecting millions of users. The source codes are released in https://github.com/onepunch-cyber/PAPERec.

Both explicit and implicit feedbacks can reflect user opinions on items, which are essential for learning user preferences in recommendation. However, most current recommendation algorithms merely focus on implicit positive feedbacks (e.g., click), ignoring other informative user behaviors. In this paper, we aim to jointly consider explicit/implicit and positive/negative feedbacks to learn user unbiased preferences for recommendation. Specifically, we propose a novel Deep feedback network (DFN) modeling click, unclick and dislike behaviors. DFN has an internal feedback interaction component that captures fine-grained interactions between individual behaviors, and an external feedback interaction component that uses precise but relatively rare feedbacks (click/dislike) to extract useful information from rich but noisy feedbacks (unclick). In experiments, we conduct both offline and online evaluations on a real-world recommendation system WeChat Top Stories used by millions of users. The significant improvements verify the effectiveness and robustness of DFN. The source code is in https://github.com/qqxiaochongqq/DFN.

The social media prediction task is aiming at predicting content popularity which includes social multimedia data such as photos, videos, and news. The task can not only help make better decisions for recommendation, but also reveals the public attention from evolutionary social systems. In this paper, we propose a novel approach named curriculum learning for wide multimedia-based transformer with graph target detection(CL-WMTG). The curriculum learning is designed for the transformer to improve the efficiency of model convergence. The mechanism of wide multimedia-based transformer is to make the model capable of learning cross information from text, pictures and other features(e.g. categories, location). Moreover, the graph target detection part can extract different features in the picture by pretrained model and reconstruct the features with a homogeneous graph network. We achieved third place in the SMP Challenge 2020.