Local deployment of MoE-based LLMs on CPU-GPU nodes is attractive for cost and privacy reasons, yet limited GPU memory forces most experts into CPU memory. Prefill is then compute-bound and benefits from aggressive GPU loading, while decode is I/O-bound and requires conservative PCIe transfers. These demands cannot be amortized under low batching---typical of local assistants and document tools---forcing phase-agnostic schedulers into a seesaw: favoring prefill's aggressive loading injects unmaskable I/O that stalls decode, while favoring decode's conservative transfers starves prefill's GPU of work. We observe that expert routing is predictable in phase-specific ways, allowing us to decouple the two phases and optimize prediction, scheduling, and execution independently for each phase. PDScope predicts expert activations to guide placement and prefetching, adapts these policies to each phase's bottleneck via cross-layer cost modeling, and compiles decisions into overhead-free pipelines using a static computation graph, cache prefetching, and zero-copy asynchronous I/O. On mainstream MoE models, PDScope achieves 2.5-10.3x higher prefill throughput and 2.1-6.1x higher decode throughput than existing systems, showing that phase-aware co-design unlocks efficient MoE inference under memory constraints and extends naturally to multi-batch settings.
Ping Luo, Jiahuan Wang, Ziqing Wen, Kele Xu, Dongsheng Li, Tao Sun
In natural science applications, federated learning (FL) enables collaborative training of models that approximate solutions to partial differential equations (PDEs) for repeated use in distributed settings. However, the input conditions of FL-PDE training samples involve multiple coupled parameters, making data heterogeneity across clients difficult to analyse. We propose solution-space PDE-Dirichlet, which groups existing solution labels into task-specific bins and distributes samples using the classical Dirichlet allocation rule in FL, providing a common partitioning approach across PDE tasks. Theoretical analysis shows an association between differences in solution-bin composition and gradient disagreement, while experiments on three controlled operators and four public simulation datasets quantify the relationships between this heterogeneity, gradient disagreement and prediction error. Based on this connection, we develop Solution-Bin Controlled Federated Optimization, which uses pooled bin-gradient prototypes to correct the component of client gradients associated with bin composition while preserving the sample-weighted training objective. Experiments on low-viscosity Burgers and Darcy flow show that this correction reduces prediction error compared with federated averaging and narrows the gap from bin-balanced IID training.
Controllable world simulation requires more than visually plausible frames: scenes must retain persistent state, coordinated actions, and editable temporal structure. Generative world models offer flexible visual synthesis, but their outputs do not always expose the objects and transitions needed for precise revision. Graphics engines provide executable state and direct control, yet translating language into valid operations requires specialized knowledge of assets, geometry, and interdependent APIs. We present CineU (Cinematize from Unreal), an agentic system for language-driven world orchestration and cinematic choreography within Unreal Engine. From a code-based world modeling perspective, an Unreal project serves as a persistent, inspectable world representation, while agent-generated code specifies operations whose effects can be executed and verified. CineU combines hierarchical orchestration, cascaded asset retrieval and geometry-constrained placement, capability-aligned code synthesis, and multi-stage verification with localized repair. Shared scene and timeline context coordinates entities, actions, camera trajectories, and shot boundaries while supporting dependency-aware revisions. Experiments on scene construction, selective editing, multi-shot cinematography, and component ablations show how specialized execution knowledge and engine feedback improve task completion and targeted recovery. CineU illustrates how executable world state can complement learned visual generation to support controllable and revisable simulated worlds.
Ziqing Wen, Ping Luo, Jiahuan Wang, Xiaoge Deng, Tao Sun, Dongsheng Li
Score-based Generative Models (SGMs) have achieved remarkable success in generation tasks by establishing a diffusion process that gradually perturbs real data into Gaussian noise and then learning a reverse process to reconstruct the data from noise. However, conventional denoising methods, such as Langevin dynamics and numerical stochastic differential equation solvers, rely on randomness but suffer from slow generation speeds due to the high number of score function evaluations. To address this, we propose a plug-and-play Markov adaptive momentum sampling method, inspired by the strong connection between the SGMs' sampling process and stochastic gradient descent optimization. Our approach accelerates the transformation process without introducing additional hyperparameters. Theoretically, we establish a rigorous foundation by analyzing the relationship between the distribution of samples generated by our method and the original data distribution. This provides a novel perspective on the analysis of sampling distributions. Empirically, our method demonstrates superior efficiency, generating high-quality images/graphs in significantly fewer sampling steps, achieving a speedup of 2x to 5x while maintaining or surpassing baseline performance on various image and graph generation benchmarks. These results underscore the effectiveness of adaptive momentum in enhancing the efficiency of SGMs.
Wanyang Hong, Zhaoning Zhang, Yi Chen, Libo Zhang, Baihui Liu, Linbo Qiao, Zhiliang Tian, Dongsheng Li
Large language models often lose track of persistent instructions and relevant information as multi-turn conversations grow. We study this cumulative contextual decay through three related failure modes: attention pollution, dilution, and drift. We propose REA(Role-aware Heuristic Episodic Attention), a context-management framework that assigns different persistence and representation policies to instructions and episodic interactions. Instructional Memory retains identified global constraints in a dedicated prefix. Episodic Memory preserves user inputs and compresses model replies, while heuristic retrieval selects raw text, compressed representations, or omission for each historical turn. On Long-MT-Bench+, REA improves the judge score from 6.32 to 7.36 on a 10-point scale, a 16.5% relative gain over the Vanilla baseline, and reduces average latency by 2.91x. Additional evaluations show aggregate gains on three backbones spanning 1.7B-7B parameters and on Chinese and English role-playing tasks. These results support role-aware context management as a practical approach to maintaining conversational continuity and instruction adherence.
Tianle Liu, Zhiliang Tian National University of Defense Technology, Zhen Huang National University of Defense Technology, Tianlun Liu National University of Defense Technology, Jingyuan Huang, Zhaoning Zhang National University of Defense Technology, Chengcheng Shao, Dongsheng Li National University of Defense Technology
As the text generated by large language models (LLMs) increasingly resembles human-written text (HWT), detecting LLM-generated text (LGT) is crucial to avoid malicious use of LGT. Recent research treats LGT detection as an out-of-distribution (OOD) detection problem and views HWT as the OOD. However, existing OOD detection methods assume that LGT is a single homogeneous distribution. In practice, LGT exhibits different characteristics under different generation conditions. Text from weaker LLMs tends to form distinct clusters and is easy to detect, whereas text from stronger models significantly overlaps with HWTs and is hard to detect. To address the issue, in this paper, we propose an LGT detection framework based on density-aware manifold learning and the construction of hybrid Mahalanobis energy. We apply density-aware manifold learning with Laplacian smoothness and density regularization in embedding space, amplifying differences between LGT and HWT. We further propose a density-adaptive hybrid Mahalanobis metric that combines global and local covariance via density weighting, enabling adaptation to the manifold-aware embedding space. Finally, based on the metric, we define the distribution energy as a measure of distribution discrepancy, and we employ energy learning and contrastive learning to separate distributions hierarchically, establishing a clear OOD decision boundary. Experiments show that our method outperforms strong baselines.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Xinyue Fang, Zhiliang Tian National University of Defense Technology, Zhen Huang National University of Defense Technology, Ziyi Pan, Zhihua Wen, Xi Wang, Quntian Fang, Dongsheng Li National University of Defense Technology
Existing LLM hallucination mitigation methods, including prompt engineering and model optimization, either hardly alter models’ internal knowledge or have poor cross-domain generalization. Contrastive decoding mitigates hallucinations by using layer-wise differences in LLMs. However, prior studies only explore transformer-based models (e.g., GPT), ignoring other effective frameworks like mixture-of-experts (MoE) models. Since MoE alters the traditional transformer architecture, we conduct empirical studies to investigate whether similar layer-wise differences exist in MoEs. Our results show that they do not exist in MoE with shared experts; nevertheless, across different MoEs, higher layers exhibit distinct expert activation patterns between factual and non-factual outputs. Building on these, we propose EAACD, an expert-aware adaptive contrast decoding that uses expert differences in MoE’s higher layers to mitigate hallucinations on QA tasks. EAACD splits high-layer experts into a higher-reliability group and several lower-reliability groups based on their confidence and consistency. It contrasts the higher-reliability group’s prediction with each lower-reliability group’s prediction to calibrate the model’s original predictions. To strengthen this contrast, EAACD amplifies hallucinations from lower-reliability experts via attention and masking to provide stronger negative references. EAACD outperforms all baselines on four datasets
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Tianlun Liu National University of Defense Technology, Zhiliang Tian National University of Defense Technology, Zhen Huang National University of Defense Technology, Xingzhi Zhou Huawei Technologies Ltd., Wanlong Yu, Tianle Liu, Feng Liu National University of Defense Technology, Dongsheng Li National University of Defense Technology
Text understanding application often suffers from domain shifts. To handle testing domains, domain adaptation (DA) is trained to adapt to a fixed and observed testing domain; a more challenging paradigm, test-time adaptation (TTA), cannot access the testing domain during training and online adapts to the testing samples during testing, where the samples are from a fixed domain. We aim to explore a more practical and underexplored scenario, continual test-time adaptation (CTTA) for text understanding, which involves a sequence of testing (unobserved) domains in testing. Current CTTA methods struggle in reducing error accumulation over domains and enhancing generalization to handle unobserved domains: 1) Noise-filtering reduces accumulated errors but discards useful information, and 2) accumulating historical domains enhances generalization, but it is hard to achieve adaptive accumulation. In this paper, we propose a CTTA-T (continual test-time adaptation for text understanding) framework adaptable to evolving target domains: CTTA-T adopts a teacher-student framework, where the teacher is equipped with domain awareness and generalization for evolving domains. To improve teacher predictions, we propose a refine-then-filter based on dropout-driven consistency, which calibrates predictions and removes unreliable guidance. For the adaptation–generalization trade-off, we construct a domain-aware teacher by dynamically accumulating cross-domain semantics via incremental PCA, which continuously tracks domain shifts. Experiments show CTTA-T excels baselines.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Baihui Liu National University of Defense Technology, Kaiyuan Tian National University of Defense Technology, Wei Wang National University of Defense Technology, Zhaoning Zhang National University of Defense Technology, Linbo Qiao, Dongsheng Li National University of Defense Technology
Mixture-of-Experts (MoE) has become a dominant architecture for scaling large language models due to their sparse activation mechanism. However, the substantial number of expert activations creates a critical latency bottleneck during inference, especially in resource-constrained deployment scenarios. Existing approaches that reduce expert activations potentially lead to severe model performance degradation. In this work, we introduce the concept of activation budget as a constraint on the number of expert activations and propose Alloc-MoE, a unified framework that optimizes budget allocation coordinately at both the layer and token levels to minimize performance degradation. At the layer level, we introduce Alloc-L, which leverages sensitivity profiling and dynamic programming to determine the optimal allocation of expert activations across layers. At the token level, we propose Alloc-T, which dynamically redistributes activations based on routing scores, optimizing budget allocation without increasing latency. Extensive experiments across multiple MoE models demonstrate that Alloc-MoE maintains model performance under a constrained activation budget. Especially, Alloc-MoE achieves 1.15\times prefill and 1.34\times decode speedups on DeepSeek-V2-Lite at half of the original budget.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Libo Zhang, Zhaoning Zhang National University of Defense Technology, Hongwanyang, Peng Qiao, Dongsheng Li National University of Defense Technology
Although speculative decoding is widely used to accelerate Vision-Language Models (VLMs) inference, it faces severe performance collapse when applied to Video Large Language Models (Vid-LLMs). The draft model typically falls into the trap of attention dilution and negative visual gain due to key-value cache explosion and context window mismatches. We observe a visual semantic internalization phenomenon in Vid-LLMs, indicating that critical visual semantics are implicitly encoded into text hidden states during deep-layer interactions, which renders raw visual inputs structurally redundant during deep inference. To address this, we propose the Sparrow framework, which first utilizes visually-aware text-anchored window attention via hidden state reuse to fully offload visual computation to the target model, and leverages intermediate-layer visual state bridging to train the draft model with semantic-rich intermediate states, thereby filtering out low-level visual noise. Additionally, a multi-token prediction strategy is introduced to bridge the training-inference distribution shift. Experiments show that Sparrow achieves an average speedup of 2.82x even with 25k visual tokens, effectively resolving the performance degradation in long sequences and offering a practical solution for real-time long video tasks.
尚无 AI 解读，展示原文摘要
cs.CL · cs.CV
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
AI 速览：提出 LongLLMLingua 提示压缩方法，旨在同时缓解长上下文场景下的计算成本高、性能下降和位置偏差三大问题。
In long context scenarios, large language models (LLMs) face three main challenges: higher computational cost, performance reduction, and position bias. Research indicates that LLM performance hinges on the density and position of key information in the input prompt. Inspired by these findings, we propose LongLLMLingua for prompt compression towards improving LLMs' perception of the key information to simultaneously address the three challenges. Our extensive evaluation across various long context scenarios demonstrates that LongLLMLingua not only enhances performance but also significantly reduces costs and latency. For instance, in the NaturalQuestions benchmark, LongLLMLingua boosts performance by up to 21.4% with around 4x fewer tokens in GPT-3.5-Turbo, leading to substantial cost savings. It achieves a 94.0% cost reduction in the LooGLE benchmark. Moreover, when compressing prompts of about 10k tokens at ratios of 2x-6x, LongLLMLingua can accelerate end-to-end latency by 1.4x-2.6x. Our code is available at https://aka.ms/LongLLMLingua.