In this dissertation, we explore intelligent methods for navigating complex search spaces through the integration of multi-agent methods, large language models (LLMs), and reinforcement learning (RL) across three key domains: graph-based path planning, multi-modal data analysis, and code related tasks of bug detection and optimization.For graph-based problems, we introduce two complementary approaches for multi-agent path finding (MAPF). Our pruning and prediction-based priority planning (P4) algorithm achieves scalability while delivering near-optimal solutions in dense environments through computational pruning and predictive planning. Building on this, we develop DREAM (Distributed Regional Efficient Agent Management), a framework that leverages LLMs to coordinate agents across subdivided regions, enabling practical large-scale MAPF operations in both offline and online scenarios. For offline scenarios, our experiments demonstrate scalable path planning for 500 agents at 50\% faster speed. For online planning, we successfully handle 8 times more agents than existing LLM-based methods. These results highlight the potential of combining classical algorithms and modern LLM reasoning to advance the frontiers of large-scale autonomous coordination.In the multi-modal domain, we propose a Reinforcement Learning guided Graph Diffusion Framework that bridges the gap between textual and visual data. By combining RL-driven graph diffusion with graph neural networks, our approach improves the extraction and understanding of complex relationships in large-scale multi-modal datasets, advancing the state-of-the-art in text-image alignment tasks. We achieved 95% and 91% accuracy for MRE and MORE datasets on multimodal relation extraction tasks.For code analysis, we propose a novel multi-LLM collaborative framework for potential bug detection. The method combines insights from multiple specialized LLM agents with hierarchical code annotations, enabling comprehensive code understanding at different abstraction levels. This approach enhances bug localization accuracy through the synthesis of diverse AI-generated insights while maintaining semantic context awareness. Our approach achieved an accuracy of 89% on standard LLM code bug detection benchmarks, demonstrating a 15% improvement over state-of-the-art methods. Additionally, we introduce an LLM-based interprocedural layout optimization method that dynamically improves code execution efficiency through intelligent transformation selection. This approach uses Large Language Models (LLMs) as the core reasoning engine, guided by our Proximal Policy Optimization (PPO) mechanism to make cross-function layout decisions. The LLMs analyze code structures and identify optimization opportunities across function boundaries that traditional compilers miss. Our case studies show the LLM progressively deepens its understanding of function relationships through iterations, identifying more key functions in the MCF benchmark over time. Combining LLM analysis with hardware performance feedback creates a powerful mechanism for data-driven layout decisions, yielding significant improvements in instruction throughput, branch prediction accuracy, and cache utilization on complex benchmarks.Through these contributions, we demonstrate how AI-driven approaches can effectively navigate and optimize complex search spaces across different domains, offering scalable and practical solutions for real-world applications.
