Jing Qiao

The orchestration of heterogeneous AI models is critical for intelligent supply chain decisionmaking, yet existing LLM-based orchestrators rely on qualitative heuristics or popularity signals, ignoring fundamental trade-offs between operational performance and multidimensional resource costs such as energy, latency, and carbon footprint. We propose SC-OPTIM, an energy-aware, data-driven inference framework that formulates model selection as a constrained optimization problem. SC-OPTIM integrates a Constrained Markov Decision Process with an online Adaptive Efficiency Frontier learned via Gaussian Process Regression, and solves a chance-constrained Mixed-Integer Quadratic Program to balance cost, lead time, service level, and resource constraints under uncertainty. Experiments on real-world supply chain tasks demonstrate that SC-OPTIM consistently improves decision accuracy across tasks, substantially increases accuracy-per-joule efficiency, reduces orchestration latency from seconds to milliseconds, and significantly lowers inference-related energy consumption, enabling efficient, scalable, and sustainable AI-driven supply chain optimization.

Recent advances in reinforcement learning for large language models have produced powerful agent frameworks that achieve strong performance on multi-turn tool use, interactive search, and complex reasoning. However, existing reinforcement learning frameworks for large language model agents face three critical limitations: difficulty in handling dynamic user interactions owing to reliance on pre-scripted queries, limited scalability across varying interaction horizons with fixed scaling schedules, and substantial reward engineering overhead requiring domain-specific manual tuning. We introduce Progressive Multi-Turn Reinforcement Learning for Dynamic User-Interactive Tool Agents, a novel framework that integrates progressive user-interactive training to overcome sparse reward signals, adaptive horizon management that monitors performance metrics and adjusts training complexity accordingly, and domain-adaptive tool orchestration that learns optimal tool selection patterns across domains. Extensive experiments on WebArena, TAU-Bench, Berkeley Function-Calling Leaderboard Version 3, BabyAI, and SciWorld demonstrate that our method achieves 28.4% success rate on WebArena and 76.3\% on TAU-Bench, substantially outperforming the baselines, such as ReAct (16.2%) and MUA-RL (24.6%), while maintaining 94.7% performance on embodied reasoning tasks and 78.9\% cross-domain performance retention. Our work establishes a unified framework for realistic user interaction training, performance-adaptive complexity scaling, and domain-flexible tool orchestration.

Deep learning models have achieved remarkable success across various domains, from computer vision and natural language processing to healthcare and autonomous systems. However, their inherent complexity and black-box nature pose significant challenges for understanding, trusting, and deploying these models in critical applications. This comprehensive survey examines the evolving landscape of explainable artificial intelligence (XAI) in deep learning, presenting a systematic analysis of current methodologies, theoretical foundations, and practical applications. We categorize existing approaches into gradientbased methods, attention mechanisms, concept-based explanations, and model-agnostic techniques, while analyzing their strengths, limitations, and computational requirements. Through extensive evaluation of over 120 research works, we identify key challenges including the trade-off between model performance and interpretability, the lack of standardized evaluation metrics, and the computational overhead of explanation generation. Our analysis reveals emerging trends such as the integration of causal reasoning, the development of inherently interpretable architectures, and the application of explainability techniques to large language models and multimodal systems. We present a unified framework for understanding different explanation paradigms and propose future research directions that address current limitations while advancing the field toward more transparent and trustworthy AI systems. This survey provides researchers and practitioners with a comprehensive understanding of XAI in deep learning, facilitating the development and deployment of interpretable models in safety-critical and regulated domains.

Meta-learning, commonly referred to as "learning to learn," has emerged as a transformative paradigm in machine learning, enabling models to rapidly adapt to new tasks with minimal training data. This comprehensive survey presents a systematic analysis of meta-learning and few-shot learning approaches, examining their theoretical foundations, algorithmic innovations, and practical applications across diverse domains. We categorize existing methods into three primary paradigms: optimizationbased approaches that learn initialization parameters, metricbased methods that learn embedding spaces for effective comparison, and model-based architectures that incorporate rapid adaptation mechanisms. Through extensive analysis of recent advances, we identify key challenges including task distribution shift, computational efficiency, and theoretical understanding of generalization bounds. Our survey reveals that while significant progress has been made in benchmark performance, practical deployment remains constrained by factors such as meta-overfitting, task diversity requirements, and computational overhead. We provide a unified framework for understanding different metalearning approaches and their relationships, analyze empirical results across standard benchmarks, and outline promising research directions including integration with foundation models, continual meta-learning, and applications in data-scarce domains. This survey serves as both a comprehensive reference for researchers and a practical guide for practitioners implementing meta-learning systems.

Neural machine translation (NMT) has revolutionized the field of natural language processing by enabling high-quality automatic translation between languages using deep neural networks. This comprehensive survey examines the evolution of machine translation from statistical methods to modern neural approaches, with particular emphasis on the transformer architecture and its variants that have dominated the field since 2017. We systematically review the fundamental architectures including encoder-decoder models, attention mechanisms, and transformer-based systems, analyzing their theoretical foundations and practical implementations. The survey explores critical challenges in multilingual NLP including low-resource translation, zero-shot learning, cross-lingual transfer, and multimodal translation. We investigate recent advances in massively multilingual models, examining architectures that can translate between hundreds of language pairs within a single model. Furthermore, we discuss the emergence of large language models in translation tasks, analyzing their capabilities and limitations compared to dedicated translation systems. The paper also addresses practical considerations including evaluation metrics, data augmentation techniques, and deployment strategies for production systems. We provide insights into current research trends including document-level translation, simultaneous translation, and neural translation with external knowledge. By synthesizing research from over 100 papers, this survey offers both theoretical foundations and practical guidance for researchers and practitioners working in neural machine translation and multilingual natural language processing.