Xiaoliang Fu, Jiaye Lin, Yangyi Fang, Chaowen Hu, Cong Qin, Zekai Shao, Binbin Zheng, Lu Pan, Ke Zeng. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Xiaoliang Fu, Jiaye Lin, Yangyi Fang, Binbin Zheng, Chaowen Hu, Zekai Shao, Cong Qin, Lu Pan, Ke Zeng, Xunliang Cai. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Multivariate time series forecasting involves predicting future values based on historical observations. However, existing approaches primarily rely on predefined single-scale patches or lack effective mechanisms for multi-scale feature fusion. These limitations hinder them from fully capturing the complex patterns inherent in time series, leading to constrained performance and insufficient generalizability. To address these challenges, we propose a novel architecture named Adaptive Weighted Mixture of Multi-Scale Expert Transformers (AdaMixT). Specifically, AdaMixT introduces various patches and leverages both General Pre-trained Models (GPM) and Domain-specific Models (DSM) for multi-scale feature extraction. To accommodate the heterogeneity of temporal features, AdaMixT incorporates a gating network that dynamically allocates weights among different experts, enabling more accurate predictions through adaptive multi-scale fusion. Comprehensive experiments on eight widely used benchmarks, including Weather, Traffic, Electricity, ILI, and four ETT datasets, consistently demonstrate the effectiveness of AdaMixT in real-world scenarios.

Multi-domain and multi-task learning enhance the efficiency and performance of industrial recommendation systems by integrating information from different domains/tasks to model user interests uniformly. However, existing methods suffer from the problem of representation entanglement, which limits the effective handling of commonality and specificity among various domains/tasks. In this paper, we propose a Prototype-guided Representation Projection (PRP) model to address this issue, which explores a novel direction of applying prototype learning to deal with complex domain/task relationships in the recommendation field. To identify inter-domain/task commonality, PRP initially uses a shared Mixture of Experts (MoE) architecture to learn representations for each sample, projecting them into a common prototype space across all domains/tasks. For domain/task specificity, specific feature extraction experts are employed, and sample representations are projected to the corresponding prototype spaces, constrained by an orthogonal loss to ensure the independence of those spaces. Moreover, PRP utilizes Optimal Transport (OT) to guide the correct representation projection within the prototype spaces, employing the linear combination of prototypes as the new sample representation. We conduct offline experiments on two open-source datasets and deploy our approach in an online system for A/B testing. Extensive experimental results consistently demonstrate that our approach outperforms existing methods.

Large Language Model (LLM)-based agents have recently shown impressive capabilities in complex reasoning and tool use via multi-step interactions with their environments. While these agents have the potential to tackle complicated tasks, their problem-solving process—agents' interaction trajectory leading to task completion—remains underexploited. These trajectories contain rich feedback that can navigate agents toward the right directions for solving problems correctly. Although prevailing approaches, such as Monte Carlo Tree Search (MCTS), can effectively balance exploration and exploitation, they ignore the interdependence among various trajectories and lack the diversity of search spaces, which leads to redundant reasoning and suboptimal outcomes. To address these challenges, we propose SE-Agent, a Self-Evolution framework that enables Agents to optimize their reasoning processes iteratively. Our approach revisits and enhances former pilot trajectories through three key operations: revision, recombination, and refinement. This evolutionary mechanism enables two critical advantages: (1) it expands the search space beyond local optima by intelligently exploring diverse solution paths guided by previous trajectories, and (2) it leverages cross-trajectory inspiration to efficiently enhance performance while mitigating the impact of suboptimal reasoning paths. Through these mechanisms, SE-Agent achieves continuous self-evolution that incrementally improves reasoning quality. We evaluate SE-Agent on SWE-bench Verified to resolve real-world GitHub issues. Experimental results across five strong LLMs show that integrating SE-Agent delivers up to 55% relative improvement, achieving state-of-the-art performance among all open-source agents on SWE-bench Verified.

The ultimate goal of code agents is to solve complex tasks autonomously. Although large language models (LLMs) have made substantial progress in code generation, real-world tasks typically demand full-fledged code repositories rather than simple scripts. Building such repositories from scratch remains a major challenge. Fortunately, GitHub hosts a vast, evolving collection of open-source repositories, which developers frequently reuse as modular components for complex tasks. Yet, existing frameworks like OpenHands and SWE-Agent still struggle to effectively leverage these valuable resources. Relying solely on README files provides insufficient guidance, and deeper exploration reveals two core obstacles: overwhelming information and tangled dependencies of repositories, both constrained by the limited context windows of current LLMs. To tackle these issues, we propose RepoMaster, an autonomous agent framework designed to explore and reuse GitHub repositories for solving complex tasks. For efficient understanding, RepoMaster constructs function-call graphs, module-dependency graphs, and hierarchical code trees to identify essential components, providing only identified core elements to the LLMs rather than the entire repository. During autonomous execution, it progressively explores related components using our exploration tools and prunes information to optimize context usage. Evaluated on the adjusted MLE-bench, RepoMaster achieves a 110\% relative boost in valid submissions over the strongest baseline OpenHands. On our newly released GitTaskBench, RepoMaster lifts the task-pass rate from 40.7% to 62.9% while reducing token usage by 95%. Our code and demonstration materials are publicly available at https://github.com/QuantaAlpha/RepoMaster.

Multipath Detection Algorithms (MDAs) are proposed to discover Internet topology in the presence of load balancing (LB). Existing methods assume uniformity in the load-balancing responses (LBR), i.e., responses from the successors of a LB router. However, we reveal that only 20% of the cases exhibit uniformity in the Internet. This finding significantly challenges the completeness of the Internet topology discovered using current MDAs. In this paper, we propose a novel system BayMuDA, that can estimate LBR distributions and calculate the minimum number of probes needed to statistically discover all nodes and links within a given hop. The validation on controlled topologies shows that BayMuDA discovers at least 85%/73% of nodes/links in ~90% of the cases. Our Internet-wide measurement results indicate that BayMuDA can discover millions of Internet nodes and links obscured by the state-of-the-art MDA algorithm, D-Miner, due to uneven responses.

Offloading machine learning models for network classification on high-throughput programmable switches is a promising technology, enabling line-speed in-network classification. Existing solutions are centralized, deploying a complete but heavy model on a single switch with limited hardware resources, causing unsatisfactory accuracy, network-wide resource wastage, and non-generic single-task classification. Therefore, we propose In-Forest-M, a general distributed multi-task in-network classification framework. Firstly, we develop a Lightweight Ensemble Generic Optional Model (LEGO), which can be transformed into base models with full functionality. Each switch only needs to deploy lightweight base models rather than complete ensemble models. The significant reduction in resource consumption allows the deployment of larger models with higher accuracy and more models that support diverse tasks. We employ a fine-grained enhancement mechanism to enhance the classification performance of base models. As traffic traverses different switches, In-Forest-M aggregates the classification results of multiple enhanced base models to improve accuracy further. Secondly, we introduce a two-phase resource-aware model allocation strategy that assigns different task-specific enhanced base models to switches under resource constraints and task requirements. To respond to dynamic traffic changes, we design an optimization-driven reinforcement learning algorithm. Moreover, we propose a lightweight update mechanism for flexible model scaling. Comprehensive experiments reveal that, compared with state-of-the-art in-network classification solutions in three real network topologies, In-Forest-M achieves increased accuracy and reduced switch rules while exhibiting great generality in multi-task classification.

Industrial multimedia recommendation systems extensively utilize cascade architectures to deliver personalized content for users, generally consisting of multiple stages like retrieval and ranking. However, retrieval models have long suffered from Sample Selection Bias (SSB) due to the distribution discrepancy between the exposed items used for model training and the candidates (almost unexposed) during inference, affecting recommendation performance. Traditional methods utilize retrieval candidates as augmented training data, indiscriminately treating unexposed data as negative samples, which leads to inaccuracies and noise. Some efforts rely on unbiased datasets, while they are costly to collect and insufficient for industrial models. In this paper, we propose a debiasing framework named DAMCAR, which introduces Domain Adaptation to mitigate SSB in Multimedia CAscade Recommendation systems. Firstly, we sample hard-to-distinguish samples from unexposed data to serve as the target domain, optimizing data quality and resource utilization. Secondly, adversarial domain adaptation is employed to generate pseudo-labels for each sample. To enhance robustness, we utilize Exponential Moving Average (EMA) to create a teacher model that supervises the generation of pseudo-labels via self-distillation. Finally, we obtain a retrieval model that maintains stable performance during inference through a hybrid training mechanism. We conduct offline experiments on two real-world datasets and deploy our approach in the retrieval model of a multimedia video recommendation system for online A/B testing. Comprehensive experimental results demonstrate the effectiveness of DAMCAR in practical applications.

Recently, Large Language Models (LLMs) have garnered significant attention in recommendation systems, improving recommendation performance through in-context learning or parameter-efficient fine-tuning. However, cross-domain generalization, i.e., model training in one scenario (source domain) but inference in another (target domain), is underexplored. In this paper, we present TLRec, a transfer learning framework aimed at enhancing LLMs for sequential recommendation tasks. TLRec specifically focuses on text inputs to mitigate the challenge of limited transferability across diverse domains, offering promising advantages over traditional recommendation models that heavily depend on unique identities (IDs) like user IDs and item IDs. Moreover, we leverage the source domain data to further enhance LLMs’ performance in the target domain. Initially, we employ powerful closed-source LLMs (e.g., GPT-4) and chain-of-thought techniques to construct instruction tuning data from the third-party scenario (source domain). Subsequently, we apply curriculum learning to fine-tune LLMs for effective knowledge injection and perform recommendations in the target domain. Experimental results demonstrate that TLRec achieves superior performance under the zero-shot and few-shot settings.

Given the high packet processing efficiency of programmable switches (e.g., P4 switches of Tbps), several works are proposed to offload the decision tree (DT) to P4 switches for in-network classification. Although the DT is suitable for the match-action paradigm in P4 switches, the range match rules used in the DT may not be supported across devices of different P4 standards. Additionally, emerging models including neural networks (NNs) and ensemble models, have shown their superior performance in networking tasks. But their sophisticated operations pose new challenges to the deployment of these models in switches. In this paper, we propose Mousikav2 to address these drawbacks successfully. First, we design a new tree model, i.e., the binary decision tree (BDT). Unlike the DT, our BDT consists of classification rules in the form of bits, which is a good fit for the standard ternary match supported by different hardware/software switches. Second, we introduce a teacher-student knowledge distillation architecture in Mousikav2, which enables the general transfer from other sophisticated models to the BDT. Through this transfer, sophisticated models are indirectly deployed in switches to avoid switch constraints. Finally, a lightweight P4 program is developed to perform classification tasks in switches with the BDT after knowledge distillation. Experiments on three networking tasks and three commodity switches show that Mousikav2 not only improves the classification accuracy by 3.27%, but also reduces the switch stage and memory usage by$2.00\times $and 28.67%, respectively. Code is available athttps://github.com/xgr19/Mousika.

A variety of model representation methods have been used in recent works to translate machine learning models into programmable switch rules to address network classification tasks at line-speed, i.e., in-network classification. These works generally deploy a complete but heavy model on a switch with limited hardware resources, causing both network-wide waste of resources and unsatisfactory accuracy. Therefore, we propose In-Forest, a general distributed in-network classification framework. Firstly, to improve accuracy with limited resources, we develop a Lightweight Ensemble Generic Optional Model (LEGO), which can be further enhanced into multiple enhanced base models with full functionality. Each switch only needs to deploy a simple base model, rather than the complete ensemble model. Thus, hardware resources required for both switches and the entire network can be significantly reduced. Secondly, as traffic traverses multiple switches, In-Forest aggregates the classification results from different enhanced base models for higher accuracy. Furthermore, we design a two-phase resource-aware model allocation strategy that assigns enhanced base models to switches under different scenarios. We use stable deep reinforcement learning to respond to dynamic traffic changes. Experimental results show that when compared to SwitchTree, Planter, and Netbeacon in two real network topologies, In-Forest can increase accuracy by up to 19.31%, while reducing the number of switch rules by 89.98%.

Decision trees (DT) have been used for high-speed networking classification on programmable switches. Most DT solutions, however, are static and cannot be deployed once the switch resource changes. In this paper, we propose Dryad to fast reprogram tree models when resource budgets change. In Dryad, we first develop a large and accurate “one-training-for-all“ DT (ODT) that can be quickly resized without computational retraining. ODTs are deployed in switches using a progressive search algorithm that searches the adaptations according to their resources. To achieve high accuracy and low packet latency, the adaptation leverages 1) innovative hard and soft pruning methods to compress the ODT rapidly with minimal performance loss; and 2) P4 scaling operations of match-action table arrangement and joint range-ternary match, which allow the switch to accommodate a larger (i.e., more accurate) ODT. Finally, an ODTCompiler is proposed to automatically convert the adapted ODT into a P4 program and then install it. Experimental results on three commodity switches under different resource scenarios show that Dryad achieves a higher classification F1-score (3.78 % higher), and completes the adaptation 161 × faster than other solutions.

Intelligent reflecting surface (IRS) is a promising technology to assist downlink information transmissions from a multi-antenna access point (AP) to a receiver. In this paper, we minimize the AP's transmit power by a joint optimization of the AP's active beamforming and the IRS's passive beamforming. Due to uncertain channel conditions, we formulate a robust power minimization problem subject to the receiver's signal-to-noise ratio (SNR) requirement and the IRS's power budget constraint. We propose a deep reinforcement learning (DRL) approach that can adapt the beamforming strategies from past experiences. To improve the learning performance, we derive a convex approximation as a lower bound on the robust problem, which is integrated with the DRL framework and thus promoting a novel optimization-driven deep deterministic policy gradient (DDPG) approach. In particular, when the DDPG algorithm generates a part of the action (e.g., passive beamforming), we can use the model-based convex approximation to optimize the other part of the action (e.g., active beamforming) efficiently. Our simulation results demonstrate that the optimization-driven DDPG algorithm can improve both the learning rate and reward significantly compared to the conventional DDPG algorithm.