Xintao Wang, Jian Yang, Weiyuan Li, Rui Xie, Jen-tse Huang, Jun Gao, Shuai Huang, Yueping Kang, Yuanli Guo, Hongwei Feng, Yanghua Xiao. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Product retrieval is the backbone of e-commerce search: for each user query, it identifies a high-recall candidate set from billions of items, laying the foundation for high-quality ranking and user experience. Despite extensive optimization for mainstream queries, existing systems still struggle with long-tail queries, especially knowledge-intensive ones. These queries exhibit diverse linguistic patterns, often lack explicit purchase intent, and require domain-specific knowledge reasoning for accurate interpretation. They also suffer from a shortage of reliable behavioral logs, which makes such queries a persistent challenge for retrieval optimization.

RAG systems consist of multiple modules to work together. However, these modules are usually separately trained. We argue that a system like RAG that incorporates multiple modules should be jointly optimized to achieve optimal performance. To demonstrate this, we design a specific pipeline called SmartRAG that includes a policy network and a retriever. The policy network can serve as 1) a decision maker that decides when to retrieve, 2) a query rewriter to generate a query most suited to the retriever and 3) an answer generator that produces the final response with/without the observations. We then propose to jointly optimize the whole system using a reinforcement learning algorithm, with the reward designed to encourage the system to achieve the highest performance with minimal retrieval cost. When jointly optimized, each module can be aware of how other modules are working and thus find the best way to work together as a complete system. Empirical results demonstrate that the jointly optimized system can achieve better performance than separately optimized counterparts.

As the data infrastructure of the 6G network, the data plane (DP) takes on the management and consumption functions of new data elements such as sensing and intelligence. Data reuse is crucial to the various data sources and large volumes of data, making the location of data storage in DP especially important. In the cloud-edge architecture of 6G DP, collaborative edge caching is proposed in the radio access network (RAN). The caching strategy for data has a significant impact on caching performance. The decision-making process for data caching policy in DP is modeled using the Stackelberg game. This approach encourages the core network and the RAN to accurately cache data at the network edge based on users' requests. The Existence of Stackelberg Equilibrium is proved. A genetic-based iterative caching policy search algorithm (GESSA) is proposed to solve the optimal joint base station caching policy to maximize the utilities of both the core network and base stations. Numerical simulations show that compared with the existing schemes without considering data collaboration and benefits for both participants, the proposed scheme can effectively improve edge caching system utilities and provide lower data provision delay and higher hit rate for UEs.

The rapid expansion of the Internet of Things (IoT) and communication technology has significantly increased the volume of complex data generated by terminal devices. This has created a demand for more efficient terminal devices with limited computing resources and battery energy. As a result, there is a need to investigate effective task offloading for many tasks with complex dependencies. This paper proposes a task offloading algorithm based on deep reinforcement learning and task offloading sequence to address the issue of dependent task offloading in end-edge-cloud collaboration scenarios. The algorithm introduces the topological structure of a directed acyclic graph (DAG) to represent task dependencies and determine task execution order based on task priorities. It uses a Markov Decision Process (MDP) to optimize and minimize latency and energy consumption for all terminal devices. Experimental results indicate that the proposed algorithm demonstrates better convergence and performance compared to baseline algorithms across various scenarios. It effectively reduces latency and energy consumption for all terminal devices, achieving cost reductions of 13.81%, 67.33%, and 81.04% compared to the three baseline algorithms, respectively. These results illustrate that our algorithm significantly outperforms the baselines under the specified conditions.

Federated Learning (FL) is increasingly adopted for training ML models, driven by its ability to preserve data privacy and reduce communication costs. However, the limited availability of wireless bandwidth necessitates efficient client selection and bandwidth allocation. This paper addresses the challenges arising from non-IID data, heterogeneous computing capabilities, and varying communication conditions. We introduce a novel data quality evaluation criterion that comprehensively takes into consideration factors including data size, local data label skew, and the Age of Data. Based on this evaluation criterion, we propose a Joint Efficient Energy-constrained Client Selection and Adaptive Bandwidth Allocation (EECS-Apt) algorithm that leverages data quality, computing capabilities and communication conditions. The experimental results indicate that while satisfying the accuracy requirement, the proposed algorithm can significantly reduce delay by up to 87.4%, 60.7% and 36.8%, respectively, compared to: 1) Joint Energy-constrained Random Client Selection and Average Bandwidth Allocation (ERCS-Avg), 2) Joint Energy-constrained Delay-based Client Selection and Adaptive Bandwidth Allocation (EDCS-Apt), and 3) Reliable and Age-sensitive Client Selection and Adaptive Bandwidth Allocation (RACS-Apt).

How to generate diverse, life-like, and unlimited long head/body sequences without any driving source? We argue that this under-investigated research problem is nontrivial at all, and has unique technical challenges behind it. Without semantic constraints from the driving sources, using the standard autoregressive model to generate infinitely long sequences would easily result in 1) out-of-distribution (OOD) issue due to the accumulated error, 2) insufficient diversity to produce natural and life-like motion sequences and 3) undesired periodic patterns along the time. To tackle the above challenges, we propose a systematic framework that marries the benefits of VQ-VAE and a novel token-level control policy trained with reinforcement learning using carefully designed reward functions. A high-level prior model can be easily injected on top to generate unlimited long and diverse sequences. Although we focus on no driving sources now, our framework can be generalized for controlled synthesis with explicit driving sources. Through comprehensive evaluations, we conclude that our proposed framework can address all the above-mentioned challenges and outperform other strong baselines very significantly.

Image composition is an important operation in image processing, but the inconsistency between foreground and background significantly degrades the quality of composite image. Image harmonization, aiming to make the foreground compatible with the background, is a promising yet challenging task. However, the lack of high-quality publicly available dataset for image harmonization greatly hinders the development of image harmonization techniques. In this work, we contribute an image harmonization dataset iHarmony4 by generating synthesized composite images based on COCO (resp., Adobe5k, Flickr, day2night) dataset, leading to our HCOCO (resp., HAdobe5k, HFlickr, Hday2night) sub-dataset. Moreover, we propose a new deep image harmonization method DoveNet using a novel domain verification discriminator, with the insight that the foreground needs to be translated to the same domain as background. Extensive experiments on our constructed dataset demonstrate the effectiveness of our proposed method. Our dataset and code are available at https://github.com/bcmi/Image_Harmonization_Datasets.

As one of the promising candidates for the next generation network, Named Data Networking (NDN) has more advantages than the TCP/IP network in areas such as mobility, content distribution and security. Although NDN is designed to defense the majority Distributed Denial of Service (DDoS) attack in the current Internet, it anticipates some new varietal DDoS attacks. A representative DDoS form is called Interest Flooding Attacks (IFA), which can be launched easily by overflowing the PIT and can do immeasurable damage to the NDN. The existing IFA detection and countermeasure methods are mainly based on the PIT abnormal state statistics. However, these methods may cause misjudgment and damage the legitimate users, especially in the case of low-rate DDoS attacks or network congestion. In this paper, we propose an IFA detection scheme based on cumulative entropy by monitoring the content request abnormal distribution and then provide the malicious prefix identification method by relative entropy theory. An Interest traceback countermeasure is also used to restrain the attacker after detection. Therefore, the proposed scheme can reduce the IFA misjudgment and protect the legitimate user, and at the same time, can avoid overreaction to normal traffic fluctuation. Simulation results reveal that our methods can effectively mitigate the IFA in NDN.

Content-Centric Networking (CCN) proposals rethink the communication model around named data. In-network caching and multipath routing are regarded as two fundamental features to distinguish the CCN from the current host-centric IP network. In this paper, we tackle the problem of joint collaborative caching and multipath routing in CCN. We achieve this with an online and offline combination caching scheme based on a local content popularity statistic results. Besides, we place the content heterogeneously along a path and resort to a caching aware dynamic multipath routing in a coordination fashion. The proposed scheme can increase the content diversity and improve the caching utility with the aim of minimizing the user access delay. Simulation experiments have been performed to evaluate the proposed scheme. Simulation results show that the proposed scheme is effective and outperforms the existing caching mechanisms in CCN.

Information-Centric Networking (ICN) proposals rethink the communication model around named data, in contrast with the host-centric transport view of TCP/IP. The content retrieval in ICN is natively receiver-driven, chunk-level based, multi-path forwarding and intrinsically coupled with in-network caching. To make the best use of in-network caching resources, NDN router should acquire the view of the cache resources nearby for intelligent forwarding. In this paper, we tackle the problem of joint multi-path forwarding and the in-network caching in ICN for the first time. An effective local popularity statistic method is firstly provided to differentiate popular and unpopular content at intermediate nodes. Then we propose a dedicated multi-path forwarding strategy to perform dynamic packet-by-packet request scheduling according the content popularity. At last, a novel cache replacement scheme is presented to improve the utilization of in-network caching. The evaluation results through ndnSIM simulator show that the proposed multi-path forwarding strategy outperforms the existing schemes in both light and heavy traffic load situations in ICN.

Content-Centric Networking (CCN) proposals rethink the communication model around named data. In-network caching is a fundamental feature to distinguish the CCN from the current host-centric IP network. Caching scheme by hash-routing is a prominent solution as it enables a joint consideration of content placement and request-to-cache routing through a hash function, and therefore, makes the cached contents visible in the domain and meanwhile, eliminates the content redundancy. In this paper, we present here a collaborative caching scheme in a CCN AS domain, which can fully exploit in-network caching by network clustering and hash-routing. Specifically, a heuristic algorithm is proposed to dynamically adjust caching capacity of each cluster based on its egress traffic while controlling the total path stretch incurred. The proposed scheme can improve the caching utility and reduce caching redundancy. Extensive experiments have been performed to evaluate the proposed scheme. Simulation results show that the proposed is effective and outperforms the existing caching mechanisms in CCN.

In this paper, we try to design a collaborative caching scheme in an edge Content Centric Networking (CCN). Specifically, we first decompose the arbitrary network into clusters based on dominating set. Then, we place the content heterogeneously within each cluster based on the content popularity results and resort to a dynamic request routing. Simulation results show that the proposed outperforms the existing caching mechanisms in CCN.