Zhiqiang Zhang

Beini Xie, Heng Chang, Ziwei Zhang et al. · tsinghua

Graph Neural Networks (GNNs) obtain tremendous success in modeling relational data. Still, they are prone to adversarial attacks, which are massive threats to applying GNNs to risk-sensitive domains. Existing defensive methods neither guarantee performance facing new data/tasks or adversarial attacks nor provide insights to understand GNN robustness from an architectural perspective. Neural Architecture Search (NAS) has the potential to solve this problem by automating GNN architecture designs. Nevertheless, current graph NAS approaches lack robust design and are vulnerable to adversarial attacks. To tackle these challenges, we propose a novel Robust Neural Architecture search framework for GNNs (G-RNA). Specifically, we design a robust search space for the message-passing mechanism by adding graph structure mask operations into the search space, which comprises various defensive operation candidates and allows us to search for defensive GNNs. Furthermore, we define a robustness metric to guide the search procedure, which helps to filter robust architectures. In this way, G-RNA helps understand GNN robustness from an architectural perspective and effectively searches for optimal adversarial robust GNNs. Extensive experimental results on benchmark datasets show that G-RNA significantly outperforms manually designed robust GNNs and vanilla graph NAS baselines by 12.1% to 23.4% under adversarial attacks.

Lei Liu, Xiaoyan Yang, Yue Shen et al.

Memory-augmented Large Language Models (LLMs) have demonstrated remarkable performance in long-term human-machine interactions, which basically relies on iterative recalling and reasoning of history to generate high-quality responses. However, such repeated recall-reason steps easily produce biased thoughts, \textit{i.e.}, inconsistent reasoning results when recalling the same history for different questions. On the contrary, humans can keep thoughts in the memory and recall them without repeated reasoning. Motivated by this human capability, we propose a novel memory mechanism called TiM (Think-in-Memory) that enables LLMs to maintain an evolved memory for storing historical thoughts along the conversation stream. The TiM framework consists of two crucial stages: (1) before generating a response, a LLM agent recalls relevant thoughts from memory, and (2) after generating a response, the LLM agent post-thinks and incorporates both historical and new thoughts to update the memory. Thus, TiM can eliminate the issue of repeated reasoning by saving the post-thinking thoughts as the history. Besides, we formulate the basic principles to organize the thoughts in memory based on the well-established operations, (\textit{i.e.}, insert, forget, and merge operations), allowing for dynamic updates and evolution of the thoughts. Furthermore, we introduce Locality-Sensitive Hashing into TiM to achieve efficient retrieval for the long-term conversations. We conduct qualitative and quantitative experiments on real-world and simulated dialogues covering a wide range of topics, demonstrating that equipping existing LLMs with TiM significantly enhances their performance in generating responses for long-term interactions.

Chunjing Gan, Binbin Hu, Bo Huang et al.

In this paper, we highlight that both conformity and risk preference matter in making fund investment decisions beyond personal interest and seek to jointly characterize these aspects in a disentangled manner. Consequently, we develop a novel M ulti-granularity Graph Disentangled Learning framework named MGDL to effectively perform intelligent matching of fund investment products. Benefiting from the well-established fund graph and the attention module, multi-granularity user representations are derived from historical behaviors to separately express personal interest, conformity and risk preference in a fine-grained way. To attain stronger disentangled representations with specific semantics, MGDL explicitly involve two self-supervised signals, i.e., fund type based contrasts and fund popularity. Extensive experiments in offline and online environments verify the effectiveness of MGDL.

Sicong Xie, Binbin Hu, Fengze Li et al.

Aiming at helping users locally discovery retail services (e.g., entertainment and dinning), Online to Offline (O2O) service platforms have become popular in recent years, which greatly challenge current recommender systems. With the real data in Alipay, a feeds-like scenario for O2O services, we find that recurrence based temporal patterns and position biases commonly exist in our scenarios, which seriously threaten the recommendation effectiveness. To this end, we propose COUPA, an industrial system targeting for characterizing user preference with following two considerations: (1) Time aware preference: we employ the continuous time aware point process equipped with an attention mechanism to fully capture temporal patterns for recommendation. (2) Position aware preference: a position selector component equipped with a position personalization module is elaborately designed to mitigate position bias in a personalized manner. Finally, we carefully implement and deploy COUPA on Alipay with a cooperation of edge, streaming and batch computing, as well as a two-stage online serving mode, to support several popular recommendation scenarios. We conduct extensive experiments to demonstrate that COUPA consistently achieves superior performance and has potential to provide intuitive evidences for recommendation

Jinluan Yang, Dingnan Jin, Anke Tang et al.

Achieving balanced alignment of large language models (LLMs) in terms of Helpfulness, Honesty, and Harmlessness (3H optimization) constitutes a cornerstone of responsible AI. Existing methods like data mixture strategies face limitations, including heavy reliance on expert knowledge and conflicting optimization signals. While model merging offers parameter-level conflict-resolution strategies through integrating specialized models' parameters, its potential for 3H optimization remains underexplored. This paper systematically compares the effectiveness of model merging and data mixture methods in constructing 3H-aligned LLMs for the first time, revealing previously overlooked collaborative and conflict relationships among the 3H dimensions and discussing the advantages and drawbacks of data mixture (\textit{data-level}) and model merging (\textit{parameter-level}) methods in mitigating the conflict for balanced 3H optimization. Specially, we propose a novel \textbf{R}eweighting \textbf{E}nhanced task \textbf{S}ingular \textbf{M}erging method, \textbf{RESM}, through outlier weighting and sparsity-aware rank selection strategies to address the challenges of preference noise accumulation and layer sparsity adaptation inherent in 3H-aligned LLM merging. Extensive evaluations can verify the effectiveness and robustness of RESM compared to previous data mixture (2\%-5\% gain) and model merging (1\%-3\% gain) methods in achieving balanced LLM alignment. We release our models through \href{https://huggingface.co/Jinluan}{3H\_Merging} for further investigations.

Wei Chen, Guo Ye, Yakun Wang et al.

Unsupervised Graph Domain Adaptation (UGDA) seeks to bridge distribution shifts between domains by transferring knowledge from labeled source graphs to given unlabeled target graphs. Existing UGDA methods primarily focus on aligning features in the latent space learned by graph neural networks (GNNs) across domains, often overlooking structural shifts, resulting in limited effectiveness when addressing structurally complex transfer scenarios. Given the sensitivity of GNNs to local structural features, even slight discrepancies between source and target graphs could lead to significant shifts in node embeddings, thereby reducing the effectiveness of knowledge transfer. To address this issue, we introduce a novel approach for UGDA called Target-Domain Structural Smoothing (TDSS). TDSS is a simple and effective method designed to perform structural smoothing directly on the target graph, thereby mitigating structural distribution shifts and ensuring the consistency of node representations. Specifically, by integrating smoothing techniques with neighborhood sampling, TDSS maintains the structural coherence of the target graph while mitigating the risk of over-smoothing. Our theoretical analysis shows that TDSS effectively reduces target risk by improving model smoothness. Empirical results on three real-world datasets demonstrate that TDSS outperforms recent state-of-the-art baselines, achieving significant improvements across six transfer scenarios. The code is available in https://github.com/cwei01/TDSS.

Zhouyu Jiang, Mengshu Sun, Zhiqiang Zhang et al.

Retrieval-Augmented Generation (RAG) effectively reduces hallucinations in Large Language Models (LLMs) but can still produce inconsistent or unsupported content. Although LLM-as-a-Judge is widely used for RAG hallucination detection due to its implementation simplicity, it faces two main challenges: the absence of comprehensive evaluation benchmarks and the lack of domain-optimized judge models. To bridge these gaps, we introduce \textbf{Bi'an}, a novel framework featuring a bilingual benchmark dataset and lightweight judge models. The dataset supports rigorous evaluation across multiple RAG scenarios, while the judge models are fine-tuned from compact open-source LLMs. Extensive experimental evaluations on Bi'anBench show our 14B model outperforms baseline models with over five times larger parameter scales and rivals state-of-the-art closed-source LLMs. We will release our data and models soon at https://github.com/OpenSPG/KAG.

Daixin Wang, Zhiqiang Zhang, Yeyu Zhao et al.

User financial default prediction plays a critical role in credit risk forecasting and management. It aims at predicting the probability that the user will fail to make the repayments in the future. Previous methods mainly extract a set of user individual features regarding his own profiles and behaviors and build a binary-classification model to make default predictions. However, these methods cannot get satisfied results, especially for users with limited information. Although recent efforts suggest that default prediction can be improved by social relations, they fail to capture the higher-order topology structure at the level of small subgraph patterns. In this paper, we fill in this gap by proposing a motif-preserving Graph Neural Network with curriculum learning (MotifGNN) to jointly learn the lower-order structures from the original graph and higherorder structures from multi-view motif-based graphs for financial default prediction. Specifically, to solve the problem of weak connectivity in motif-based graphs, we design the motif-based gating mechanism. It utilizes the information learned from the original graph with good connectivity to strengthen the learning of the higher-order structure. And considering that the motif patterns of different samples are highly unbalanced, we propose a curriculum learning mechanism on the whole learning process to more focus on the samples with uncommon motif distributions. Extensive experiments on one public dataset and two industrial datasets all demonstrate the effectiveness of our proposed method.

Ling Team, Bin Hu, Cai Chen et al.

We present Ring-lite, a Mixture-of-Experts (MoE)-based large language model optimized via reinforcement learning (RL) to achieve efficient and robust reasoning capabilities. Built upon the publicly available Ling-lite model, a 16.8 billion parameter model with 2.75 billion activated parameters, our approach matches the performance of state-of-the-art (SOTA) small-scale reasoning models on challenging benchmarks (e.g., AIME, LiveCodeBench, GPQA-Diamond) while activating only one-third of the parameters required by comparable models. To accomplish this, we introduce a joint training pipeline integrating distillation with RL, revealing undocumented challenges in MoE RL training. First, we identify optimization instability during RL training, and we propose Constrained Contextual Computation Policy Optimization(C3PO), a novel approach that enhances training stability and improves computational throughput via algorithm-system co-design methodology. Second, we empirically demonstrate that selecting distillation checkpoints based on entropy loss for RL training, rather than validation metrics, yields superior performance-efficiency trade-offs in subsequent RL training. Finally, we develop a two-stage training paradigm to harmonize multi-domain data integration, addressing domain conflicts that arise in training with mixed dataset. We will release the model, dataset, and code.

Dingyuan Zhu, Daixin Wang, Zhiqiang Zhang et al.

Uplift modeling aims to measure the incremental effect, which we call uplift, of a strategy or action on the users from randomized experiments or observational data. Most existing uplift methods only use individual data, which are usually not informative enough to capture the unobserved and complex hidden factors regarding the uplift. Furthermore, uplift modeling scenario usually has scarce labeled data, especially for the treatment group, which also poses a great challenge for model training. Considering that the neighbors' features and the social relationships are very informative to characterize a user's uplift, we propose a graph neural network-based framework with two uplift estimators, called GNUM, to learn from the social graph for uplift estimation. Specifically, we design the first estimator based on a class-transformed target. The estimator is general for all types of outcomes, and is able to comprehensively model the treatment and control group data together to approach the uplift. When the outcome is discrete, we further design the other uplift estimator based on our defined partial labels, which is able to utilize more labeled data from both the treatment and control groups, to further alleviate the label scarcity problem. Comprehensive experiments on a public dataset and two industrial datasets show a superior performance of our proposed framework over state-of-the-art methods under various evaluation metrics. The proposed algorithms have been deployed online to serve real-world uplift estimation scenarios.

Xiong Jun Wu, Zhenduo Zhang, ZuJie Wen et al.

Training large reasoning models (LRMs) with reinforcement learning in STEM domains is hindered by the scarcity of high-quality, diverse, and verifiable problem sets. Existing synthesis methods, such as Chain-of-Thought prompting, often generate oversimplified or uncheckable data, limiting model advancement on complex tasks. To address these challenges, we introduce SHARP, a unified approach to Synthesizing High-quality Aligned Reasoning Problems for LRMs reinforcement learning with verifiable rewards (RLVR). SHARP encompasses a strategic set of self-alignment principles -- targeting graduate and Olympiad-level difficulty, rigorous logical consistency, and unambiguous, verifiable answers -- and a structured three-phase framework (Alignment, Instantiation, Inference) that ensures thematic diversity and fine-grained control over problem generation. We implement SHARP by leveraging a state-of-the-art LRM to infer and verify challenging STEM questions, then employ a reinforcement learning loop to refine the model's reasoning through verifiable reward signals. Experiments on benchmarks such as GPQA demonstrate that SHARP-augmented training substantially outperforms existing methods, markedly improving complex reasoning accuracy and pushing LRM performance closer to expert-level proficiency. Our contributions include the SHARP strategy, framework design, end-to-end implementation, and experimental evaluation of its effectiveness in elevating LRM reasoning capabilities.

Jingwei Wang, Zai Zhang, Hao Qian et al.

Teaching large language models (LLMs) to use tools is crucial for improving their problem-solving abilities and expanding their applications. However, effectively using tools is challenging because it requires a deep understanding of tool functionalities and user intentions. Previous methods relied mainly on LLMs to generate instruction data, but the quality of these data was often insufficient. In this paper, we propose a new method that uses knowledge graphs to generate high-quality instruction data for LLMs. Knowledge graphs are manually curated datasets rich in semantic information. We begin by extracting various query pathways from a given knowledge graph, which are transformed into a broad spectrum of user queries. We then translate the relationships between entities into actionable tools and parse the pathways of each query into detailed solution steps, thereby creating high-quality instruction data. Our experiments show that fine-tuning on just a small sample of this synthetic data can significantly improve the tool utilization and overall capabilities of LLMs.

Junze Chen, Cheng Yang, Shujie Li et al.

Large language models (LLMs) have demonstrated their strong capabilities in various domains, and have been recently integrated for graph analysis as graph language models (GLMs). With LLMs as the predictor, some GLMs can interpret unseen tasks described by natural language, and learn from a few examples in the prompts without parameter tuning, known as in-context learning (ICL). Another subset of GLMs utilizes abundant training labels to enhance model performance, known as instruction tuning. However, we argue that ICL on graphs has effectiveness issues due to fixed parameters and efficiency issues due to long context. Meanwhile, the large amount of labeled data required for instruction tuning can be difficult to obtain in real-world scenarios. To this end, we aim to introduce an extra parameter adaptation stage that can efficiently tailor GLMs to an unseen graph and task with only a few labeled examples, in exchange for better prediction accuracy and faster inference speed. For implementation, in this paper we propose GraphLAMA method, with its model backbone and learning schemes specialized for efficient tuning and inference. Specifically, for model backbone, we use a graph neural network (GNN) with several well-designed components to transform nodes into the representation space of LLM tokens. Task instructions can then be represented as a mixture of node and language tokens. In the pre-training stage, model parameters except the LLM will be trained with different tasks to capture general knowledge. In the adaptation stage, only a few pre-trained parameters will be updated based on few-shot examples. Extensive experiments on few/zero-shot node classification and summary generation show that our proposed GraphLAMA achieves state-of-the-art performance with 4.91% absolution improvement in accuracy. Compared with ICL, our inference speed can be 10 times faster under 5-shot setting.

Kai Zhang, Hao Qian, Qi Liu et al.

Recent studies in recommender systems have managed to achieve significantly improved performance by leveraging reviews for rating prediction. However, despite being extensively studied, these methods still suffer from some limitations. First, previous studies either encode the document or extract latent sentiment via neural networks, which are difficult to interpret the sentiment of reviewers intuitively. Second, they neglect the personalized interaction of reviews with user/item, i.e., each review has different contributions when modeling the sentiment preference of user/item. To remedy these issues, we propose a Sentiment-aware Interactive Fusion Network (SIFN) for review-based item recommendation. Specifically, we first encode user/item reviews via BERT and propose a light-weighted sentiment learner to extract semantic features of each review. Then, we propose a sentiment prediction task that guides the sentiment learner to extract sentiment-aware features via explicit sentiment labels. Finally, we design a rating prediction task that contains a rating learner with an interactive and fusion module to fuse the identity (i.e., user and item ID) and each review representation so that various interactive features can synergistically influence the final rating score. Experimental results on five real-world datasets demonstrate that the proposed model is superior to state-of-the-art models.

Jinghan Shi, Houye Ji, Chuan Shi et al.

The prosperous development of e-commerce has spawned diverse recommendation systems. As a matter of fact, there exist rich and complex interactions among various types of nodes in real-world recommendation systems, which can be constructed as heterogeneous graphs. How learn representative node embedding is the basis and core of the personalized recommendation system. Meta-path is a widely used structure to capture the semantics beneath such interactions and show potential ability in improving node embedding. In this paper, we propose Heterogeneous Graph neural network for Recommendation (HGRec) which injects high-order semantic into node embedding via aggregating multi-hops meta-path based neighbors and fuses rich semantics via multiple meta-paths based on attention mechanism to get comprehensive node embedding. Experimental results demonstrate the importance of rich high-order semantics and also show the potentially good interpretability of HGRec.

Dalong Zhang, Xianzheng Song, Ziqi Liu et al.

Link prediction is widely used in a variety of industrial applications, such as merchant recommendation, fraudulent transaction detection, and so on. However, it's a great challenge to train and deploy a link prediction model on industrial-scale graphs with billions of nodes and edges. In this work, we present a scalable and distributed framework for semi-supervised link prediction problem (named DSSLP), which is able to handle industrial-scale graphs. Instead of training model on the whole graph, DSSLP is proposed to train on the \emph{$k$-hops neighborhood} of nodes in a mini-batch setting, which helps reduce the scale of the input graph and distribute the training procedure. In order to generate negative examples effectively, DSSLP contains a distributed batched runtime sampling module. It implements uniform and dynamic sampling approaches, and is able to adaptively construct positive and negative examples to guide the training process. Moreover, DSSLP proposes a model-split strategy to accelerate the speed of inference process of the link prediction task. Experimental results demonstrate that the effectiveness and efficiency of DSSLP in serval public datasets as well as real-world datasets of industrial-scale graphs.