华为

The advent of large language models (LLMs), such as GPT, Gemini, and DeepSeek, has significantly advanced natural language processing, giving rise to sophisticated chatbots capable of diverse language-related tasks. The transition from these traditional LLM chatbots to more advanced AI agents represents a pivotal evolutionary step. However, existing evaluation frameworks often blur the distinctions between LLM chatbots and AI agents, leading to confusion among researchers selecting appropriate benchmarks. To bridge this gap, this paper introduces a systematic analysis of current evaluation approaches, grounded in an evolutionary perspective. We provide a detailed analytical framework that clearly differentiates AI agents from LLM chatbots along five key aspects: complex environment, multi-source instructor, dynamic feedback, multi-modal perception, and advanced capability. Further, we categorize existing evaluation benchmarks based on external environments driving forces, and resulting advanced internal capabilities. For each category, we delineate relevant evaluation attributes, presented comprehensively in practical reference tables. Finally, we synthesize current trends and outline future evaluation methodologies through four critical lenses: environment, agent, evaluator, and metrics. Our findings offer actionable guidance for researchers, facilitating the informed selection and application of benchmarks in AI agent evaluation, thus fostering continued advancement in this rapidly evolving research domain.

The optimal operation of transportation systems is often susceptible to unexpected disruptions. Many established control strategies reliant on mathematical models can struggle with real-world disruptions, leading to significant divergence from their anticipated efficiency. This study integrates the cutting-edge concept of antifragility with learning-based traffic control strategies to optimize urban road network operations under disruptions. Antifragile systems not only withstand and recover from stressors but also thrive and enhance performance in the presence of such adverse events. We develop a deep reinforcement learning framework that incorporates antifragile modules composed of traffic state derivatives and redundancy. Subsequently, it is evaluated in a cordon-shaped transportation network and a case study with real-world data. Promising results highlight that the proposed algorithm provides: (i) superior performance achieving up to 27.6% and 41.9% performance gain over baselines under increasing demand and supply disruptions, (ii) lower distribution skewness under disruptions, demonstrating its relative antifragility against baselines, (iii) effectiveness under limited observability due to real-world data availability constraints, and (iv) the robustness and transferability to be incorporated into recent state-of-the-art RL architectures. The proposed antifragile methodology is generalizable and holds potential for applications beyond traffic engineering, offering integration into control systems exposed to disruptions across various disciplines.

Cross-modal remote sensing image-text retrieval (CMRSITR) plays a critical role in extracting semantic information from Earth observation data. However, applying parameter-efficient prompt tuning to this task faces a critical causal chain of failures: (1) the limited trainable parameters induce representation instability on noisy remote sensing data, (2) this instability hinders the bridging of the persistent modality gap, and (3) the resulting misalignment renders traditional sampling strategies ineffective, leading to the mining of false negatives. To break this dependency chain, we introduce Progressive Representation Stabilization and Alignment (ProSA), a novel framework that enforces a stabilize-align-discriminate curriculum. ProSA first employs a Hierarchical Uncertainty Modeling (HUM) module to inject adaptive noise, smoothing the optimization landscape to establish representation stability. Building upon this stable foundation, an Adversarial Modality Alignment (AMA) module is introduced to bridge the modality gap via a gradient reversal mechanism. Finally, only within the stabilized and aligned feature space, a Hard Negative Mining (HNM) module is activated to refine fine-grained discriminability by targeting valid hard negatives. Extensive experiments on three benchmark datasets demonstrate that ProSA effectively resolves these challenges, achieving state-of-the-art performance among parameter-efficient methods and competitive results against computation-heavy full fine-tuning baselines.

The rise of large models has made distributed computing networks indispensable for training, as the computational and memory demands of these models far exceed the capabilities of any single computing device. Frequent data transmission and parameter synchronization among computing devices within computing networks make network reliability a critical factor that directly impacts both operational stability and model training efficiency. However, existing research lacks a precise definition and quantitative analysis methodology specifically tailored for computing networks, particularly concerning the coupling effects between dynamic service and underlying physical infrastructure. To address this issue, this paper proposes a computing network model oriented to logical-to-physical network mapping. The model incorporates both the static structure and the dynamic temporal states of the network, thereby accounting for the intricate mapping relationships between the logical service processes and the physical infrastructure. Subsequently, a dual-layer, Bayesian network that incorporates both static and dynamic elements is constructed, enabling a quantitative reliability analysis based on the network model. To validate the proposed method, a case study on distributed training of the GPT-3 model is conducted. Comparative reliability experiments on typical architectures show that the Torus structure achieves higher reliability, offering foundational insights for designing robust computing networks.