Large language model (LLM) agents face fundamental limitations in long-horizon reasoning due to finite context windows, making effective memory management critical.Existing methods typically handle long-term memory (LTM) and short-term memory (STM) as separate components, relying on heuristics or auxiliary controllers, which limits adaptability and end-to-end optimization.In this paper, we propose Agentic Memory (AgeMem), a unified framework that integrates LTM and STM management directly into the agent's policy.AgeMem exposes memory operations as tool-based actions, enabling the LLM agent to autonomously decide what and when to store, retrieve, update, summarize, or discard information.To train such unified behaviors, we propose a three-stage progressive reinforcement learning strategy and design a step-wise GRPO to address sparse and discontinuous rewards induced by memory operations.Experiments on five long-horizon benchmarks demonstrate that AgeMem consistently outperforms strong memory-augmented baselines across multiple LLM backbones, achieving improved task performance, higher-quality long-term memory, and more efficient context usage.

Multimodal retrieval has emerged as a promising yet challenging research direction in recent years. Most existing studies in multimodal retrieval focus on capturing information in multimodal data that is similar to their paired texts, but often ignores the complementary information contained in multimodal data. In this study, we propose CIEA, a novel multimodal retrieval approach that employs Complementary Information Extraction and Alignment, which transforms both text and images in documents into a unified latent space and features a complementary information extractor designed to identify and preserve differences in the image representations. We optimize CIEA using two complementary contrastive losses to ensure semantic integrity and effectively capture the complementary information contained in images. Extensive experiments demonstrate the effectiveness of CIEA, which achieves significant improvements over both divide-and-conquer models and universal dense retrieval models. We provide an ablation study, further discussions, and case studies to highlight the advancements achieved by CIEA. To promote further research in the community, we have released the source code at https://github.com/zengdlong/CIEA.

Despite significant progress in abstractive text summarization aimed at generating fluent and informative outputs, how to ensure the factual consistency of generated summaries remains a crucial and challenging issue. In this study, drawing inspiration from advancements in causal inference, we construct causal graphs to analyze the process of abstractive text summarization methods and identify intrinsic causes of factual inconsistency, specifically language bias and irrelevancy bias, and we propose CoFactSum, a novel framework that mitigates the causal effects of these biases through counterfactual estimation for enhancing the factual consistency of the generated content. CoFactSum provides two counterfactual estimation strategies, including Explicit Counterfactual Masking, which employs a dynamic masking approach, and Implicit Counterfactual Training, which utilizes a discriminative cross-attention mechanism. Besides, we propose a Debiasing Degree Adjustment mechanism to dynamically calibrate the level of debiasing at each decoding step. Extensive experiments conducted on two widely used summarization datasets demonstrate the effectiveness and advantages of the proposed CoFactSum in enhancing the factual consistency of generated summaries, outperforming several baseline methods.

There has been a surge of interest regarding language adaptation of Large Language Models (LLMs) to enhance the processing of texts in low-resource languages. While traditional language models have seen extensive research on language transfer, modern LLMs still necessitate further explorations in language adaptation. In this paper, we present a systematic review of the language adaptation process for LLMs, including vocabulary expansion, continued pre-training, and instruction fine-tuning, which focuses on empirical studies conducted on LLaMA2 and discussions on various settings affecting the model’s capabilities. This study provides helpful insights covering the entire language adaptation process, and highlights the compatibility and interactions between different steps, offering researchers a practical guidebook to facilitate the effective adaptation of LLMs across different languages.

Through the collaboration of multiple LLM-empowered agents possessing diverse expertise and tools, multi-agent systems achieve impressive progress in solving real-world problems. Given the user queries, the meta-agents, serving as the brain within multi-agent systems, are required to decompose the queries into multiple sub-tasks that can be allocated to suitable agents capable of solving them, so-called agent-oriented planning. In this study, we identify three critical design principles of agent-oriented planning, including solvability, completeness, and non-redundancy, to ensure that each sub-task can be effectively resolved, resulting in satisfactory responses to user queries. These principles further inspire us to propose AOP, a novel framework for agent-oriented planning in multi-agent systems, leveraging a fast task decomposition and allocation process followed by an effective and efficient evaluation via a reward model. According to the evaluation results, the meta-agent is also responsible for promptly making necessary adjustments to sub-tasks and scheduling. Besides, we integrate a feedback loop into AOP to further enhance the effectiveness and robustness of such a problem-solving process. Extensive experiments demonstrate the advancement of AOP in solving real-world problems compared to both single-agent systems and existing planning strategies for multi-agent systems. The source code is available at https://github.com/lalaliat/Agent-Oriented-Planning

This study addresses the challenge of noise in training datasets for Direct Preference Optimization (DPO), a method for aligning Large Language Models (LLMs) with human preferences. We categorize noise into pointwise noise, which includes low-quality data points, and pairwise noise, which encompasses erroneous data pair associations that affect preference rankings. Utilizing Distributionally Robust Optimization (DRO), we enhance DPO's resilience to these types of noise. Our theoretical insights reveal that DPO inherently embeds DRO principles, conferring robustness to pointwise noise, with the regularization coefficient $\beta$ playing a critical role in its noise resistance. Extending this framework, we introduce Distributionally Robustifying DPO (Dr. DPO), which integrates pairwise robustness by optimizing against worst-case pairwise scenarios. The novel hyperparameter $\beta'$ in Dr. DPO allows for fine-tuned control over data pair reliability, providing a strategic balance between exploration and exploitation in noisy training environments. Empirical evaluations demonstrate that Dr. DPO substantially improves the quality of generated text and response accuracy in preference datasets, showcasing enhanced performance in both noisy and noise-free settings.

Federated learning (FL) has emerged as a promising paradigm for fine-tuning foundation models using distributed data in a privacy-preserving manner. Under limited computational resources, clients often find it more practical to fine-tune a selected subset of layers, rather than the entire model, based on their task-specific data. In this study, we provide a thorough theoretical exploration of selective layer fine-tuning in FL, emphasizing a flexible approach that allows the clients to adjust their selected layers according to their local data and resources. We theoretically demonstrate that the layer selection strategy has a significant impact on model convergence in two critical aspects: the importance of selected layers and the heterogeneous choices across clients. Drawing from these insights, we further propose a strategic layer selection method that utilizes local gradients and regulates layer selections across clients. Extensive experiments on both image and text datasets demonstrate the effectiveness of the proposed strategy compared with several baselines, highlighting its advances in identifying critical layers that adapt to the client heterogeneity and training dynamics in FL.

Xiangrong Zhu, Yuexiang Xie, Yi Liu, Yaliang Li, Wei Hu. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

Text-to-image diffusion models have gained widespread application across various domains, demonstrating remarkable creative potential. However, the strong generalization capabilities of diffusion models can inadvertently lead to the generation of not-safe-for-work (NSFW) content, posing significant risks to their safe deployment. While several concept erasure methods have been proposed to mitigate the issue associated with NSFW content, a comprehensive evaluation of their effectiveness across various scenarios remains absent. To bridge this gap, we introduce a full-pipeline toolkit specifically designed for concept erasure and conduct the first systematic study of NSFW concept erasure methods. By examining the interplay between the underlying mechanisms and empirical observations, we provide in-depth insights and practical guidance for the effective application of concept erasure methods in various real-world scenarios, with the aim of advancing the understanding of content safety in diffusion models and establishing a solid foundation for future research and development in this critical area.

Large language models (LLMs) have demonstrated great capabilities in various natural language understanding and generation tasks.These pre-trained LLMs can be further improved for specific downstream tasks by fine-tuning.However, the adoption of LLM in real-world applications can be hindered by privacy concerns and the resource-intensive nature of model training and fine-tuning.When multiple entities have similar interested tasks but cannot directly share their local data due to privacy regulations, federated learning (FL) is a mainstream solution to leverage the data of different entities.Besides avoiding direct data sharing, FL can also achieve rigorous data privacy protection, model intelligent property protection, and model customization via composition with different techniques.Despite the aforementioned advantages of FL, fine-tuning LLMs in FL settings still lacks adequate support from the existing frameworks and, therefore, faces challenges in optimizing the consumption of significant communication and computational resources, preparing various data for different tasks, and satisfying diverse information protection demands.

Nowadays, as privacy concerns continue to rise, federated graph learning (FGL) which generalizes the classic federated learning to graph data has attracted increasing attention. However, while the focus has been on designing collaborative learning algorithms, the potential risks of privacy leakage through the sharing of necessary graph-related information in FGL, such as node embeddings and neighbor generators, have been largely neglected. In this paper, we verify the potential risks of privacy leakage in FGL, and provide insights about the cautions in FGL algorithm design. Specifically, we propose a novel privacy attack algorithm named Privacy Attack on federated Graph learning (PAG) towards reconstructing participants’ private node attributes and the linkage relationships. The participant performing the PAG attack is able to reconstruct the node attributes of the victim by matching the received gradients of the generator, and then train a link prediction model based on its local sub-graph to inductively infer the linkages connected to these reconstructed nodes. We theoretically and empirically demonstrate that under PAG attack, directly sharing the neighbor generators makes the FGL vulnerable to the data reconstruction attack. Furthermore, an investigation into the key factors that can hinder the success of the PAG attack provides insights into corresponding defense strategies and inspires future research into privacy-preserving FGL.

As people inevitably interact with items across multiple domains or various platforms, cross-domain recommendation (CDR) has gained increasing attention. However, the rising privacy concerns limit the practical applications of existing CDR models, since they assume that full or partial data are accessible among different domains. Recent studies on privacy-aware CDR models neglect the heterogeneity from multiple-domain data and fail to achieve consistent improvements in cross-domain recommendation; thus, it remains a challenging task to conduct effective CDR in a privacy-preserving way. In this article, we propose a novel, as far as we know, federated graph learning approach for Privacy-Preserving Cross-Domain Recommendation (PPCDR) to capture users’ preferences based on distributed multi-domain data and improve recommendation performance for all domains without privacy leakage. The main idea of PPCDR is to model both global preference among multiple domains and local preference at a specific domain for a given user, which characterizes the user’s shared and domain-specific tastes toward the items for interaction. Specifically, in the private update process of PPCDR, we design a graph transfer module for each domain to fuse global and local user preferences and update them based on local domain data. In the federated update process, through applying the local differential privacy technique for privacy-preserving, we collaboratively learn global user preferences based on multi-domain data and adapt these global preferences to heterogeneous domain data via personalized aggregation. In this way, PPCDR can effectively approximate the multi-domain training process that directly shares local interaction data in a privacy-preserving way. Extensive experiments on three CDR datasets demonstrate that PPCDR consistently outperforms competitive single- and cross-domain baselines and effectively protects domain privacy.

Based on the assumption that samples in the source and target domains are freely accessible during training, unsupervised domain adaptation (UDA) of question answering (QA) aims to transfer knowledge learned from labeled source datasets to similar tasks in the unlabeled target domains. However, such assumption can easily lead to privacy violation issues in real-world applications, especially when the source domain data involves privacy-intensive domains such as finance and healthcare. In this paper, we introduce Source-Free Domain Adaptation Framework for QA (denoted as SFQA), which only allows access to trained source models for target learning, making data privacy protection more promising. Specifically, the proposed SFQA model consists of a feature extractor module (Bert Encoder) and a classifier module (Answer Classifier). We first transfer the trained source model to the target model while keeping the source classifier module frozen. Then we adopt the question generation model to generate questions and answers for the target domain. Taking the generated questions and target domain context as inputs, and the generated answers as pseudo-labels, we train the target model with joint entropy to learn a target domain-specific feature extractor. Experimental results demonstrate the superiority and effectiveness of the proposed SFQA, and show that SFQA outperforms the state-of-the-art methods.

Federated Learning (FL) aims to train high-quality models in collaboration with distributed clients while not uploading their local data, which attracts increasing attention in both academia and industry. However, there is still a considerable gap between the flourishing FL research and real-world scenarios, mainly caused by the characteristics of heterogeneous devices and its scales. Most existing works conduct evaluations with homogeneous devices, which are mismatched with the diversity and variability of heterogeneous devices in real-world scenarios. Moreover, it is challenging to conduct research and development at scale with heterogeneous devices due to limited resources and complex software stacks. These two key factors are important yet underexplored in FL research as they directly impact the FL training dynamics and final performance, making the effectiveness and usability of FL algorithms unclear. To bridge the gap, in this paper, we propose an efficient and scalable prototyping system for real-world cross-device FL, FS-REAL. It supports heterogeneous device runtime, contains parallelism and robustness enhanced FL server, and provides implementations and extensibility for advanced FL utility features such as personalization, communication compression and asynchronous aggregation. To demonstrate the usability and efficiency of FS-REAL, we conduct extensive experiments with various device distributions, quantify and analyze the effect of the heterogeneous device and various scales, and further provide insights and open discussions about real-world FL scenarios. Our system is released to help to pave the way for further real-world FL research and broad applications involving diverse devices and scales.

Federated learning (FL) is a general distributed machine learning paradigm that provides solutions for tasks where data cannot be shared directly. Due to the difficulties in communication management and heterogeneity of distributed data and devices, initiating and using an FL algorithm for real-world cross-device scenarios requires significant repetitive effort but may not be transferable to similar projects. To reduce the effort required for developing and deploying FL algorithms, we present FS-Real, an open-source FL platform designed to address the need of a general and efficient infrastructure for real-world cross-device FL. In this paper, we introduce the key components of FS-Real and demonstrate that FS-Real has the following capabilities: 1) reducing the programming burden of FL algorithm development with plug-and-play and adaptable runtimes on Android and other Internet of Things (IoT) devices; 2) handling a large number of heterogeneous devices efficiently and robustly with our communication management components; 3) supporting a wide range of advanced FL algorithms with flexible configuration and extension; 4) alleviating the costs and efforts for deployment, evaluation, simulation, and performance optimization of FL algorithms with automatized tool kits.

Although remarkable progress has been made by existing federated learning (FL) platforms to provide infrastructures for development, these platforms may not well tackle the challenges brought by various types of heterogeneity. To fill this gap, in this paper, we propose a novel FL platform, named FederatedScope, which employs an event-driven architecture to provide users with great flexibility to independently describe the behaviors of different participants. Such a design makes it easy for users to describe participants with various local training processes, learning goals and backends, and coordinate them into an FL course with synchronous or asynchronous training strategies. Towards an easy-to-use and flexible platform, FederatedScope enables rich types of plug-in operations and components for efficient further development, and we have implemented several important components to better help users with privacy protection, attack simulation and auto-tuning. We have released FederatedScope at https://github.com/alibaba/FederatedScope to promote academic research and industrial deployment of federated learning in a wide range of scenarios.

The lottery ticket hypothesis (LTH) states that a randomly initialized dense network contains sub-networks that can be trained in isolation to the performance of the dense network. In this paper, to achieve rapid learning with less computational cost, we explore LTH in the context of meta learning. First, we experimentally show that there are sparse sub-networks, known as meta winning tickets, which can be meta-trained to few-shot classification accuracy to the original backbone. The application of LTH in meta learning enables the adaptation of meta-trained networks on various IoT devices with fewer computation. However, the status quo to identify winning tickets requires iterative training and pruning, which is particularly expensive for finding meta winning tickets. To this end, then we investigate the inter- and intra-layer patterns among different meta winning tickets, and propose a scheme for early detection of a meta winning ticket. The proposed scheme enables efficient training in resource-limited devices. Besides, it also designs a lightweight solution to search the meta winning ticket. Evaluations on standard few-shot classification benchmarks show that we can find competitive meta winning tickets with 20% weights of the original backbone, while incurring only 8%-14% (Conv-4) and 19%-29% (ResNet-12) computation overhead (measured by FLOPs) of the standard winning ticket finding scheme.

The incredible development of federated learning (FL) has benefited various tasks in the domains of computer vision and natural language processing, and the existing frameworks such as TFF and FATE has made the deployment easy in real-world applications. However, federated graph learning (FGL), even though graph data are prevalent, has not been well supported due to its unique characteristics and requirements. The lack of FGL-related framework increases the efforts for accomplishing reproducible research and deploying in real-world applications. Motivated by such strong demand, in this paper, we first discuss the challenges in creating an easy-to-use FGL package and accordingly present our implemented package FederatedScope-GNN (FS-G), which provides (1) a unified view for modularizing and expressing FGL algorithms; (2) comprehensive DataZoo and ModelZoo for out-of-the-box FGL capability; (3) an efficient model auto-tuning component; and (4) off-the-shelf privacy attack and defense abilities. We validate the effectiveness of FS-G by conducting extensive experiments, which simultaneously gains many valuable insights about FGL for the community. Moreover, we employ FS-G to serve the FGL application in real-world E-commerce scenarios, where the attained improvements indicate great potential business benefits. We publicly release FS-G, as submodules of FederatedScope, at https://github.com/alibaba/FederatedScope to promote FGL's research and enable broad applications that would otherwise be infeasible due to the lack of a dedicated package.

Recently, graph neural networks (GNN) have been successfully applied to recommender systems as an effective collaborative filtering (CF) approach. However, existing GNN-based CF models suffer from noisy user-item interaction data, which seriously affects the effectiveness and robustness in real-world applications. Although there have been several studies on data denoising in recommender systems, they either neglect direct intervention of noisy interaction in the message-propagation of GNN, or fail to preserve the diversity of recommendation when denoising.

Although neural networks have achieved great successes in various machine learning tasks, people can hardly know what neural networks learn from data due to their black-box nature. The lack of such explainability is one of the limitations of neural networks when applied in domains, e.g., healthcare and finance, that demand transparency and accountability. Moreover, explainability is beneficial for guiding a neural network to learn the causal patterns that can extrapolate out-of-distribution (OOD) data, which is critical in real-world applications and has surged as a hot research topic.

Answer selection, which is involved in many natural language processing applications, such as dialog systems and question answering (QA), is an important yet challenging task in practice, since conventional methods typically suffer from the issues of ignoring diverse real-world background knowledge. In this article, we extensively investigate approaches to enhancing the answer selection model with external knowledge from knowledge graph (KG). First, we present a context-knowledge interaction learning framework, Knowledge-aware Neural Network, which learns the QA sentence representations by considering a tight interaction with the external knowledge from KG and the textual information. Then, we develop two kinds of knowledge-aware attention mechanism to summarize both the context-based and knowledge-based interactions between questions and answers. To handle the diversity and complexity of KG information, we further propose a Contextualized Knowledge-aware Attentive Neural Network, which improves the knowledge representation learning with structure information via a customized Graph Convolutional Network and comprehensively learns context-based and knowledge-based sentence representation via the multi-view knowledge-aware attention mechanism. We evaluate our method on four widely used benchmark QA datasets, including WikiQA, TREC QA, InsuranceQA, and Yahoo QA. Results verify the benefits of incorporating external knowledge from KG and show the robust superiority and extensive applicability of our method.

Machine Learning methods have been adopted for a wide range of real-world applications, ranging from social networks, online image/video-sharing platforms, and e-commerce to education, healthcare, etc. However, in practice, a large amount of effort is required to tune several components of machine learning methods, including data representation, hyperparameter, and model architecture, in order to achieve a good performance. To alleviate the required tunning efforts, Automated Machine Learning (AutoML), which can automate the process of applying machine learning methods, has been studied in both academy and industry recently. In this tutorial, we will introduce the main research topics of AutoML, including Hyperparameter Optimization, Neural Architecture Search, and Meta-Learning. Two emerging topics of AutoML, Automatic Feature Generation and Machine Learning Guided Database, will also be discussed since they are important components for real-world applications. For each topic, we will motivate it with application examples from industry, illustrate the state-of-the-art methodologies, and discuss some future research directions based on our experience from industry and the trends in academy.

High-order interactive features capture the correlation between different columns and thus are promising to enhance various learning tasks on ubiquitous tabular data. To automate the generation of interactive features, existing works either explicitly traverse the feature space or implicitly express the interactions via intermediate activations of some designed models. These two kinds of methods show that there is essentially a trade-off between feature interpretability and search efficiency. To possess both of their merits, we propose a novel method named Feature Interaction Via Edge Search (FIVES), which formulates the task of interactive feature generation as searching for edges on the defined feature graph. Specifically, we first present our theoretical evidence that motivates us to search for useful interactive features with increasing order. Then we instantiate this search strategy by optimizing both a dedicated graph neural network (GNN) and the adjacency tensor associated with the defined feature graph. In this way, the proposed FIVES method simplifies the time-consuming traversal as a typical training course of GNN and enables explicit feature generation according to the learned adjacency tensor. Experimental results on both benchmark and real-world datasets show the advantages of FIVES over several state-of-the-art methods. Moreover, the interactive features identified by FIVES are deployed on the recommender system of Taobao, a worldwide leading e-commerce platform. Results of an online A/B testing further verify the effectiveness of the proposed method FIVES, and we further provide FIVES as AI utilities for the customers of Alibaba Cloud.

Community question answering (CQA) gains increasing popularity in both academy and industry recently. However, the redundancy and lengthiness issues of crowdsourced answers limit the performance of answer selection and lead to reading difficulties and misunderstandings for community users. To solve these problems, we tackle the tasks of answer selection and answer summary generation in CQA with a novel joint learning model. Specifically, we design a question-driven pointer-generator network, which exploits the correlation information between question-answer pairs to aid in attending the essential information when generating answer summaries. Meanwhile, we leverage the answer summaries to alleviate noise in original lengthy answers when ranking the relevancy degrees of question-answer pairs. In addition, we construct a new large-scale CQA corpus, WikiHowQA, which contains long answers for answer selection as well as reference summaries for answer summarization. The experimental results show that the joint learning method can effectively address the answer redundancy issue in CQA and achieves state-of-the-art results on both answer selection and text summarization tasks. Furthermore, the proposed model is shown to be of great transferring ability and applicability for resource-poor CQA tasks, which lack of reference answer summaries.

We study the community question answering (CQA) problem that emerges with the advent of numerous community forums in the recent past. The task of finding appropriate answers to questions from informative but noisy crowdsourced answers is important yet challenging in practice. We present an Attentive User-engaged Adversarial Neural Network (AUANN), which interactively learns the context information of questions and answers, and enhances user engagement with the CQA task. A novel attentive mechanism is incorporated to model the semantic internal and external relations among questions, answers and user contexts. To handle the noise issue caused by introducing user context, we design a two-step denoise mechanism, including a coarse-grained selection process by similarity measurement, and a fine-grained selection process by applying an adversarial training module. We evaluate the proposed method on large-scale real-world datasets SemEval-2016 and SemEval-2017. Experimental results verify the benefits of incorporating user information, and show that our proposed model significantly outperforms the state-of-the-art methods.

Answer selection and knowledge base question answering (KBQA) are two important tasks of question answering (QA) systems. Existing methods solve these two tasks separately, which requires large number of repetitive work and neglects the rich correlation information between tasks. In this paper, we tackle answer selection and KBQA tasks simultaneously via multi-task learning (MTL), motivated by the following motivations. First, both answer selection and KBQA can be regarded as a ranking problem, with one at text-level while the other at knowledge-level. Second, these two tasks can benefit each other: answer selection can incorporate the external knowledge from knowledge base (KB), while KBQA can be improved by learning contextual information from answer selection. To fulfill the goal of jointly learning these two tasks, we propose a novel multi-task learning scheme that utilizes multi-view attention learned from various perspectives to enable these tasks to interact with each other as well as learn more comprehensive sentence representations. The experiments conducted on several real-world datasets demonstrate the effectiveness of the proposed method, and the performance of answer selection and KBQA is improved. Also, the multi-view attention scheme is proved to be effective in assembling attentive information from different representational perspectives.