Large reasoning models (LRMs) have achieved remarkable performance in complex reasoning tasks, driven by their powerful inference-time scaling capability.However, LRMs often suffer from overthinking, which results in substantial computational redundancy and significantly reduces efficiency.Early-exit methods aim to mitigate this issue by terminating reasoning once sufficient evidence has been generated, yet existing approaches mostly rely on handcrafted or empirical indicators that are unreliable and impractical.In this work, we introduce Dynamic Thought Sufficiency in Reasoning (DTSR), a novel framework for efficient reasoning that enables the model to dynamically assess the sufficiency of its chain-of-thought (CoT) and determine the optimal point for early exit.Inspired by human metacognition, DTSR operates in two stages: (1) Reflection Signal Monitoring, which identifies reflection signals as potential cues for early exit, and (2) Thought Sufficiency Check, which evaluates whether the current CoT is sufficient to derive the final answer.Experimental results on the Qwen3 models show that DTSR reduces reasoning length by 28.9%-34.9%with minimal performance loss, effectively mitigating overthinking.We further discuss overconfidence in LRMs and self-evaluation paradigms, providing valuable insights for early-exit reasoning.

With the rapid growth of video-based applications, edge-cloud collaboration has become a mainstream paradigm for large-scale visual inference. However, existing edge-cloud systems primarily emphasize task offloading and static resource allocation, often overlooking the dynamic and heterogeneous nature of real-world scenarios. The significant variability in scene complexity across tasks leads to inefficient system performance. In this article, we propose CLAP, a cross-layer adaptive pipelining inference scheduling framework for edge-cloud vision systems. First, CLAP introduces a lightweight multiscale scene-aware module that accurately characterizes the visual complexity of incoming tasks at different granularities with minimal overhead. Based on this complexity profile, we design an adaptive multi-stage pipeline scheduling strategy, which dynamically adjusts processing granularity and selectively activates stages across edge and cloud nodes. Furthermore, we formulate the resource allocation as a multi-agent decision-making problem and employ cross-layer reinforcement learning to optimize task distribution under complex objectives, efficiently balancing accuracy, delay, and energy consumption. Extensive evaluations on public datasets demonstrate that CLAP can improve the throughput by more than 2.1x compared to traditional cloud-only and edge-only solutions while meeting accuracy requirements. Compared to state-of-the-art edge-cloud methods, CLAP achieves a 3% improvement in inference accuracy while simultaneously reducing end-to-end resource overhead, delay, and energy consumption by over 35%, proving its effectiveness in dynamic, large-scale vision applications.

Heterogeneous device-edge-cloud computing infrastructures have become the backbone of modern telecommunication operators and Wide Area Networks (WANs), providing multi-tier computational support for emerging intelligent applications. With the rapid proliferation of Large Language Model (LLM) services, efficiently coordinating inference tasks and reducing communication burden within these multi-tier network architectures becomes a critical deployment challenge. Current LLM serving paradigms exhibit significant limitations: on-device deployment restricts service to lightweight LLMs due to hardware constraints, while cloud-centric deployment encounters resource congestion and considerable prompt communication overhead during peak periods. Model-cascading inference, though better suited for multi-tier networks, depends on static, manually-tuned thresholds that cannot adapt to dynamic network conditions or varying task complexities. To address these challenges, we propose RecServe, a recursive offloading framework tailored for LLM serving in multi-tier networks. RecServe introduces a task-specific hierarchical confidence evaluation mechanism that guides offloading decisions based on inferred task complexity in progressively scaled LLMs across device, edge, and cloud tiers. To further enable intelligent task routing across tiers, RecServe employs a sliding-window-based dynamic offloading strategy with quantile interpolation, enabling real-time tracking of historical confidence distributions and adaptive offloading threshold adjustments. This design allows inference tasks to be recursively offloaded to higher tiers only when necessary, optimizing heterogeneous resource utilization while reducing cross-tier communication with little compromise on service quality. Theoretical analysis provides distinct conditions under which RecServe is expected to achieve reduced communication burden and computational costs. Experiments on eight datasets demonstrate that RecServe outperforms CasServe in both service quality and communication efficiency, and reduces the communication burden by over 50% compared to centralized cloud-based serving.

Large Language Models (LLMs), with remarkable conversational capabilities, have emerged as AI assistants that can handle both visual and textual modalities. However, their effectiveness in joint video-language understanding has not been extensively explored. In the article, we introduce Valley , a multi-modal foundation model designed to enable enhanced video comprehension and instruction-following capabilities. To this end, we construct two datasets, namely “ Valley-702k ” and “ Valley-instruct-73k ,” to cover a diverse range of video–text alignment and video-based instruction tasks, such as multi-shot captions, long video descriptions, action recognition, causal inference, and so on. Then, we adopt ViT-L/14 as the vision encoder and explore three different temporal modeling modules to learn multifaceted features for enhanced video understanding. In addition, we implement a two-phase training approach for Valley: the first phase focuses solely on training the projection module to facilitate the LLM’s capacity to understand visual input, and the second phase jointly trains the projection module and the LLM to improve their instruction following ability. Extensive experiments demonstrate that Valley has the potential to serve as an effective video assistant, simplifying complex video understanding tasks. Our code and data are publicly available at https://github.com/RupertLuo/Valley .

Scaling multi-dimensional transformers to long sequences is indispensable across various domains. However, the challenges of large memory requirements and slow speeds of such sequences necessitate sequence parallelism. All existing approaches fall under the category of embedded sequence parallelism, which are limited to shard along a single sequence dimension, thereby introducing significant communication overhead. However, the nature of multi-dimensional transformers involves independent calculations across multiple sequence dimensions. To this end, we propose Dynamic Sequence Parallelism (DSP) as a novel abstraction of sequence parallelism. DSP dynamically switches the parallel dimension among all sequences according to the computation stage with efficient resharding strategy. DSP offers significant reductions in communication costs, adaptability across modules, and ease of implementation with minimal constraints. Experimental evaluations demonstrate DSP's superiority over state-of-the-art embedded sequence parallelism methods by remarkable throughput improvements ranging from 32.2% to 10x, with less than 25% communication volume.

Cloud inference consumes massive amounts of resources and carbon emissions, which has prompted edge-cloud inference to become a new paradigm. However, constrained by limited edge resources, existing approaches predominantly adopt periodic model switching strategies, overlooking the heterogeneity of application requirements and the highly dynamic nature of task demands in large-scale edge-cloud environments. To address these challenges, this paper proposes an Adaptive Online Dynamic Model Switching (AODMS) framework, which jointly optimizes real-time model switching, cross-layer resource allocation, and edge-cloud collaborative inference to minimize long-term system cost, including task execution time and carbon emissions. First, an average saving cost evaluation mechanism is developed to determine optimal model switching points, coupled with a matrix-constrained pairwise search strategy for efficient exploration of feasible edge model loading. And then a distributed iterative optimization scheme is proposed for joint resource allocation and edge-cloud task scheduling. Extensive experimental evaluations demonstrate that AODMS can significantly reduce model switching frequency, average inference time, and Carbon Footprint (CF), offering a scalable and adaptive solution for the sustainable edge-cloud system.

Recent multimodal large language models (MLLMs) marry modality-specific vision or audio encoders with a shared text decoder. While the encoder is compute- intensive but memory-light, the decoder is the opposite, yet state-of-the-art serving stacks still time-multiplex these complementary kernels, idling SMs or HBM in turn. We introduce SpaceServe, a serving system that space-multiplexes MLLMs: it decouples all modality encoders from the decoder, and co-locates them on the same GPU using fine-grained SM partitioning available in modern runtimes. A cost-model-guided Space-Inference Scheduler (SIS) dynamically assigns SM slices, while a Time-Windowed Shortest-Remaining-First (TWSRFT) policy batches en- coder requests to minimise completion latency and smooth decoder arrivals. Evaluation shows that SpaceServe reduces time-per-output-token by 4.81× on average and up to 28.9× on Nvidia A100 GPUs. SpaceServe is available at https://github.com/gofreelee/SpaceServe

With the development of artificial intelligence, many visual detection methods have been applied to power systems. However, diverse tasks cause existing solutions to face low detection accuracy and resource constraints in edge power scenarios. In this paper, we propose a low-rank enhanced lightweight multimodal LLM framework for efficient edge power visual detection. The framework can effectively balance the model performance with the resource constraints of edge deployment by introducing multimodal LLM and fusing data augmentation strategies and low-rank optimization. First, to enhance the adaptability of diverse visual detection tasks, we design a diverse data augmentation strategy for multimodal LLM to solve the problem of insufficient power scene data. Then, we reduce the number of training parameters by a low-rank optimization technique to enable the model to run efficiently on resource-limited edge devices. Experimental results demonstrate that our proposed framework can improve the accuracy by more than 3% on average under different power visual detection tasks. It can also save 76.49% of graphics memory consumption and accelerate training by 81.88%.

Currently, massive video inference tasks are processed through edge-cloud collaboration. However, the diverse scenarios make it difficult to allocate the inference tasks efficiently, resulting in many wasted resources. In this paper, we propose a joint-aware video processing (JAVP) architecture for edge-cloud collaboration. First, we develop a multiscale complexity-aware model for predicting task complexity and determining its suitability for edge or cloud servers. The task is subsequently efficiently scheduled to the appropriate servers by integrating complexity with an adaptive resource-aware optimization algorithm. For input tasks, JAVP can dynamically and intelligently select the most appropriate server. The evaluation results on public datasets show that JAVP can improve the through-put by more than 70% compared to traditional cloud-only solutions while meeting accuracy requirements. And JAVP can improve the accuracy by 3%-5% and reduce delay and energy consumption by 16%-50% compared to state-of-the-art edge-cloud solutions.

In visual intelligent scenarios, large amounts of real-time video data are generated at the end. During the optimization process for different video tasks, frequent data copying between devices and hosts can be limited by data bandwidth, resulting in high system latency. We investigate computing bottlenecks in online video processing to reduce processing latency and improve efficiency. In this paper, we propose a joint video acceleration processing (JVAP) architecture for online edge systems. First, video-compressed streams are transmitted to the GPU for decoding and conversion of data content. Second, we design data pre-processing and post-processing modules to achieve specific functional operators and separately complete operator combinations and stitching. Different computing tasks can reuse the implemented operator library. Third, we modify the data interface of the inference task model to maintain the consistent flow of data in the GPU. We conduct experiments using videos of different qualities and model frameworks of varying scales. The results indicate that the proposed method enhances the average processing efficiency over 11% with respect to existing representative acceleration frameworks and extends the potential application of online intelligent inference algorithms.

Recently, the Internet of Things (IoT) has played a powerful role in healthcare. However, the rapid growth of healthcare devices has produced many heterogeneous data and most of them are visual. It brings great difficulties to the calculation, cache, and transmission of data. The geographical dispersion and the dynamicity of nodes also challenge the development of healthcare IoT (HIoT). In this article, we propose an intelligent end–edge–cloud architecture for visual IoT-assisted healthcare systems (intelligent V-HIoT) to improve the end-to-end performance of next-generation smart healthcare. First, we systemically analyze the characteristics of human–machine–things in end side from the perspective of data processing, then define the end intelligence, solving the problem of intelligence measurement of heterogeneous devices. Second, we propose an efficiency intelligence measurement model in the edge side and cloud side, which provides a theoretical basis for the dynamic management of edge nodes. Third, we present an end–edge–cloud framework that optimizes the efficiency of data processing and node deployment. The intelligence level of HIoT is maximized as well as intelligent management of nodes is implemented. To verify the effectiveness, we perform the experiments for different approaches. The simulation results demonstrate that the intelligent V-HIoT significantly outperforms existing approaches because the proposed method can achieve maximum intelligence level of both in many heterogeneous devices and an emergency medical situation.