Large language models (LLMs) have shown impressive capabilities in natural language tasks, yet they continue to struggle with multi-step mathematical reasoning, where correctness depends on a precise chain of intermediate steps. Preference optimization methods such as Direct Preference Optimization (DPO) have improved answer-level alignment, but they often overlook the reasoning process itself, providing little supervision over intermediate steps that are critical for complex problem-solving. Existing fine-grained approaches typically rely on strong annotators or reward models to assess the quality of individual steps. However, reward models are vulnerable to reward hacking. To address this, we propose ISLA, a reward-model-free framework that constructs step-level preference data directly from SFT gold traces. ISLA also introduces a self-improving pruning mechanism that identifies informative steps based on two signals: their marginal contribution to final accuracy (relative accuracy) and the model’s uncertainty, inspired by the concept of information gain. Empirically, ISLA achieves better performance than DPO while using only 12% of the training tokens, demonstrating that careful step-level selection can significantly improve both reasoning accuracy and training efficiency.

Zexu Sun, Yongcheng Zeng, Erxue Min, Heyang Gao, Bokai Ji, Dugang Liu, Xing Tang, Xiuqiang He, Xu Chen. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Jiongnan Liu, Yutao Zhu, Shuting Wang, Xiaochi Wei, Erxue Min, Yu Lu, Shuaiqiang Wang, Dawei Yin, Zhicheng Dou. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Retrieval-augmented generation (RAG) has demonstrated strong potential in enhancing large language models (LLMs) for complex, real-world question answering. However, existing RAG frameworks remain inadequate for temporal scenarios, primarily due to their inability to jointly model temporal constraints in both retrieval and reasoning. On the retrieval side, traditional approaches focus on semantic similarity, often returning outdated or temporally misaligned evidence. On the generation side, these systems frequently produce factually incorrect or hallucinated answers when confronted with incomplete or temporally inconsistent information. Motivated by the observed limitations, we propose ChronoReflect+, a temporal logic-aware RAG framework that incorporates hybrid temporal-aware retrieval and progressive multi-step reflection. Our method iteratively refines both retrieval and reasoning, identifying and bridging information gaps as context accumulates. Extensive experiments demonstrate that ChronoReflect+ significantly outperforms state-of-the-art RAG baselines-improving end-to-end accuracy by 15.2%-particularly on questions involving implicit time expressions and multi-hop reasoning.

Recent advancements in LLM alignment leverage token-level supervisions to perform finegrained preference optimization.However, existing token-level alignment methods either optimize on all available tokens, which can be noisy and inefficient, or perform selective training with complex and expensive key token selection strategies.In this work, we propose Selective Preference Optimization (SePO), a novel selective alignment strategy that centers on efficient key token selection without requiring strong, fine-grained supervision signals.We prove the feasibility of Direct Preference Optimization (DPO) as token-level reward function estimators, which applies to any existing alignment datasets and enables costefficient token selection with small-scale model sizes and training data.We then train an oracle model with DPO on the target data and utilize the estimated reward function to score all tokens within the target dataset, where only the key tokens are selected to supervise the target policy model with a contrastive objective function.Extensive experiments on three public evaluation benchmarks show that SePO significantly outperforms competitive baseline methods by only optimizing on 30% key tokens with up to 60% reduction in GPU training hours.We also explore SePO as a new paradigm for weakto-strong generalization, showing that weak oracle models effectively supervise strong policy models with up to 16.8× more parameters.SePO also selects useful supervision signals from out-of-distribution data, alleviating the over-optimization problem.The project is open-sourced here.

In this work, we propose FoRA-UA, a novel method that, using only 1-5% of the standard LoRA's parameters, achieves state-ofthe-art performance across a wide range of tasks.Specifically, we explore scenarios with extremely limited parameter budgets and derive two key insights: (1) fix-sized sparse frequency representations approximate small matrices more accurately; and (2) with a fixed number of trainable parameters, introducing a smaller intermediate representation to approximate larger matrices results in lower construction error.These findings form the foundation of our FoRA-UA method.By inserting a small intermediate parameter set, we achieve greater model compression without sacrificing performance.We evaluate FoRA-UA across diverse tasks, including natural language understanding (NLU), natural language generation (NLG), instruction tuning, and image classification, demonstrating strong generalisation and robustness under extreme compression. 1

Benefiting from the strong reasoning capabilities, Large language models (LLMs) have demonstrated remarkable performance in recommender systems. Various efforts have been made to distill knowledge from LLMs to enhance collaborative models, employing techniques like contrastive learning for representation alignment. In this work, we prove that directly aligning the representations of LLMs and collaborative models is suboptimal for enhancing downstream recommendation tasks performance, based on the information theorem. Consequently, the challenge of effectively aligning semantic representations between collaborative models and LLMs remains unresolved. Inspired by this viewpoint, we propose a novel plug-and-play alignment framework for LLMs and collaborative models. Specifically, we first disentangle the latent representations of both LLMs and collaborative models into specific and shared components via projection layers and representation regularization. Subsequently, we perform both global and local structure alignment on the shared representations to facilitate knowledge transfer. Additionally, we theoretically prove that the specific and shared representations contain more pertinent and less irrelevant information, which can enhance the effectiveness of downstream recommendation tasks. Extensive experimental results on benchmark datasets demonstrate that our method is superior to existing state-of-the-art algorithms.

Recommender systems are increasingly spreading to different areas like e-commerce or video streaming to alleviate information overload. One of the most fundamental methods for recommendation is Collaborative Filtering (CF), which leverages historical user-item interactions to infer user preferences. In recent years, Graph Neural Networks (GNNs) have been extensively studied to capture graph structures in CF tasks. Despite this remarkable progress, local structure modeling and embedding distortion still remain two notable limitations in the majority of GNN-based CF methods. Therefore, in this paper, we propose a novel Hyperbolic Graph Transformer architecture, to tackle the long-tail problems in CF tasks. Specifically, the proposed framework is comprised of two essential modules: 1) Local Hyperbolic Graph Convolutional Network (LHGCN), which performs graph convolution entirely in the hyperbolic manifold and captures the local structure of each node; 2) Hyperbolic Transformer, which is comprised of hyperbolic cross-attention mechanisms to capture global information. Furthermore, to enable its feasibility on large-scale data, we introduce an unbiased approximation of the cross-attention for linear computational complexity, with a theoretical guarantee in approximation errors. Empirical experiments demonstrate that our proposed model outperforms the leading collaborative filtering methods and significantly mitigates the long-tail issue in CF tasks. Our implementations are available in https://github.com/EnkiXin/Hgformer.

Cross-Domain Recommendation (CDR) seeks to utilize knowledge from different domains to alleviate the problem of data sparsity in the target recommendation domain, and has been gaining more attention in recent years. Although there have been notable advances in this area, most current methods represent users and items in Euclidean space, which is not ideal for handling long-tail distributed data in recommendation systems. Additionally, adding data from other domains can worsen the long-tail characteristics of the entire dataset, making it harder to train CDR models effectively. Recent studies have shown that hyperbolic methods are particularly suitable for modeling long-tail distributions, which has led us to explore hyperbolic representations for users and items in CDR scenarios. However, due to the distinct characteristics of the different domains, applying hyperbolic representation learning to CDR tasks is quite challenging. In this paper, we introduce a new framework called Hyperbolic Contrastive Learning (HCTS), designed to capture the unique features of each domain while enabling efficient knowledge transfer between domains. We achieve this by embedding users and items from each domain separately and mapping them onto distinct hyperbolic manifolds with adjustable curvatures for prediction. To improve the representations of users and items in the target domain, we develop a hyperbolic contrastive learning module for knowledge transfer. Extensive experiments on real-world datasets demonstrate that hyperbolic manifolds are a promising alternative to Euclidean space for CDR tasks. The codes are available at https://github.com/EnkiXin/hcts.

Contrastive deep graph clustering (CDGC) leverages the power of contrastive learning to group nodes into different clusters. The quality of contrastive samples is crucial for achieving better performance, making augmentation techniques a key factor in the process. However, the augmentation samples in existing methods are always predefined by human experiences, and agnostic from the downstream task clustering, thus leading to high human resource costs and poor performance. To overcome these limitations, we propose a Graph Node Clustering with Fully Learnable Augmentation, termed GraphLearner. It introduces learnable augmentors to generate high-quality and task-specific augmented samples for CDGC. GraphLearner incorporates two learnable augmentors specifically designed for capturing attribute and structural information. Moreover, we introduce two refinement matrices, including the high-confidence pseudo-label matrix and the cross-view sample similarity matrix, to enhance the reliability of the learned affinity matrix. During the training procedure, we notice the distinct optimization goals for training learnable augmentors and contrastive learning networks. In other words, we should both guarantee the consistency of the embeddings as well as the diversity of the augmented samples. To address this challenge, we propose an adversarial learning mechanism within our method. Besides, we leverage a two-stage training strategy to refine the high-confidence matrices. Extensive experimental results on six benchmark datasets validate the effectiveness of GraphLearner.The code and appendix of GraphLearner are available at https://github.com/xihongyang1999/GraphLearner on Github.

Traditional Click-Through Rate (CTR) prediction models are usually trained and deployed in a single scenario. However, large-scale commercial platforms usually contain multiple recommendation scenarios, the traffic characteristics of which may be significantly different. Recent studies have proved that learning a unified model to serve multiple scenarios is effective in improving the overall performance. However, most existing approaches suffer from various limitations respectively, such as insufficient distinction modeling, inefficiency with the increase of scenarios, and lack of interpretability. More importantly, as far as we know, none of existing Multi-Scenario Modeling approaches takes explicit feature interaction into consideration when modeling scenario distinctions, which limits the expressive power of the network and thus impairs the performance. In this paper, we propose a novel Scenario-Adaptive Feature Interaction framework named SATrans, which models scenario discrepancy as the distinction of patterns in feature correlations. Specifically, SATrans is built on a Transformer architecture to learn high-order feature interaction and involves the scenario information in the modeling of self-attention to capture distribution shifts across scenarios. We provide various implementations of our framework to boost the performance, and experiments on both public and industrial datasets show that SATrans 1) significantly outperforms existing state-of-the-art approaches for prediction, 2) is parameter-efficient as the space complexity grows marginally with the increase of scenarios, 3) offers good interpretability in both instance-level and scenario-level. We have deployed the model in WeChat Official Account Platform and have seen more than 2.84% online CTR increase on average in three major scenarios.

Click-Through Rate (CTR) prediction, which aims to estimate the probability that a user will click an item, is an essential component of online advertising. Existing methods mainly attempt to mine user interests from users' historical behaviours, which contain users' directly interacted items. Although these methods have made great progress, they are often limited by the recommender system's direct exposure and inactive interactions, and thus fail to mine all potential user interests. To tackle these problems, we propose Neighbor-Interaction based CTR prediction (NI-CTR), which considers this task under a Heterogeneous Information Network (HIN) setting. In short, Neighbor-Interaction based CTR prediction involves the local neighborhood of the target user-item pair in the HIN to predict their linkage. In order to guide the representation learning of the local neighbourhood, we further consider different kinds of interactions among the local neighborhood nodes from both explicit and implicit perspective, and propose a novel Graph-Masked Transformer (GMT) to effectively incorporates these kinds of interactions to produce highly representative embeddings for the target user-item pair. Moreover, in order to improve model robustness against neighbour sampling, we enforce a consistency regularization loss over the neighbourhood embedding. We conduct extensive experiments on two real-world datasets with millions of instances and the experimental results show that our proposed method outperforms state-of-the-art CTR models significantly. Meanwhile, the comprehensive ablation studies verify the effectiveness of every component of our model. Furthermore, we have deployed this framework on the WeChat Official Account Platform with billions of users. The online A/B tests demonstrate an average CTR improvement of 21.9% against all online baselines.

Fake News detection has attracted much attention in recent years. Social context based detection methods attempt to model the spreading patterns of fake news by utilizing the collective wisdom from users on social media. This task is challenging for three reasons: (1) There are multiple types of entities and relations in social context, requiring methods to effectively model the heterogeneity. (2) The emergence of news in novel topics in social media causes distribution shifts, which can significantly degrade the performance of fake news detectors. (3) Existing fake news datasets usually lack of great scale, topic diversity and user social relations, impeding the development of this field. To solve these problems, we formulate social context based fake news detection as a heterogeneous graph classification problem, and propose a fake news detection model named Post-User Interaction Network (PSIN), which adopts a divide-and-conquer strategy to model the post-post, user-user and post-user interactions in social context effectively while maintaining their intrinsic characteristics. Moreover,we adopt an adversarial topic discriminator for topic-agnostic feature learning, in order to improve the generalizability of our method for new-emerging topics. Furthermore, we curate a new dataset for fake news detection, which contains over 27,155 news from 5 topics, 5 million posts, 2 million users and their induced social graph with 0.2 billion edges. It has been published on https://github.com/qwerfdsaplking/MC-Fake. Extensive experiments illustrate that our method outperforms SOTA baselines in both in-topic and out-of-topic settings.

As we head towards the IoT (Internet of Things) era, protecting network infrastructures and information security has become increasingly crucial. In recent years, Anomaly-Based Network Intrusion Detection Systems (ANIDSs) have gained extensive attention for their capability of detecting novel attacks. However, most ANIDSs focus on packet header information and omit the valuable information in payloads, despite the fact that payload-based attacks have become ubiquitous. In this paper, we propose a novel intrusion detection system named TR-IDS, which takes advantage of both statistical features and payload features. Word embedding and text-convolutional neural network (Text-CNN) are applied to extract effective information from payloads. After that, the sophisticated random forest algorithm is performed on the combination of statistical features and payload features. Extensive experimental evaluations demonstrate the effectiveness of the proposed methods.

Stochastic gradient descent (SGD) is a commonly used technique in large-scale machine learning tasks, but its convergence is slow due to the inherent variance. In recent years, a popular method, Stochastic Variance Reduced Gradient (SVRG), addresses this shortcoming via computing the full gradient of the entire dataset in each epoch. However, conventional SVRG and its variants usually need to identify a hyperparameter - the epoch size, which is essential to the convergence performance. Few previous studies discuss how to systematically find a suitable value for that hyper-parameter, which makes it hard to gain a good convergence performance in practical machine learning applications. In this paper, we propose a new stochastic gradient descent named AESVRG, which introduces variance reduction and computes the full gradient adaptively. Its enhanced implementation, AESVRG+, has a convergence performance that can outplay existing SVRG with fine-tuned epoch sizes. An extensive evaluation illustrates the significant performance improvement of our method.