Addressing the core challenge of balancing cost and efficiency in probe deployment within blockchain peer-to-peer (P2P) networks, traditional static or simple heuristic strategies often struggle with the high dynamics, node heterogeneity, and frequent topological fluctuations, leading to resource waste or monitoring blind spots. To achieve more intelligent and adaptive network monitoring, this study innovatively proposes a dynamic probe deployment framework based on Large Language Models (LLMs). This framework aims to fully leverage the powerful information processing, pattern recognition, and reasoning capabilities of LLMs to build an integrated “Perception-Cognition-Decision” closed-loop intelligent system, with the goal of achieving the most efficient network state information acquisition using optimized resource consumption (minimizing probe count). The core architecture of the framework comprises three collaborative modules: The Perception Layer transforms heterogeneous network data into structured spatio-temporal embeddings, utilizing dynamic graph neural networks to capture topological dependencies and temporal evolution. The Cognition Layer relies on the LLM for network state prediction (e.g., trends in node importance, regional density changes, and uncertainty quantification) and adaptive parameter generation (e.g., importance score weights, monitoring radius factors), enabling data-driven strategy adaptation. The Decision Layer integrates current state, future predictions, and dynamic parameters, evaluates node value based on the LLM, and performs constrained optimization (e.g., probe count limit K) to generate the optimal deployment plan that maximizes coverage and minimizes transmission delay. These three modules work synergistically, forming a real-time responsive and intelligent decision-making system for network dynamics. To preliminarily validate the potential and feasibility of this LLM framework, this research designed and implemented a specific probe deployment optimization algorithm as an instance of the framework's decision logic. This algorithm focuses on a three-dimensional node importance assessment (type, connectivity, stability), combined with dynamic monitoring radius adjustment based on regional density, and employs a greedy strategy to prioritize the selection of highest-value nodes. Extensive experimental results on simulated P2P networks (power-law model) demonstrate that this specific algorithm, compared to the baseline full deployment strategy, can significantly save approximately 66% of probe resources while ensuring 100% network coverage. Furthermore, it effectively reduces the Information Collection Cost (ICC) by about 33% compared to a random deployment strategy. The experiments also confirm the algorithm's good scalability, with deployment costs showing a controllable linear growth relative to network size, and demonstrate intelligent adaptability to network changes across different scales. In summary, this research not only proposes a forward-looking LLM-based dynamic probe deployment framework, offering new intelligent approaches to tackle blockchain P2P network monitoring challenges, but also successfully validates the effectiveness of a specific optimization algorithm within this framework. The demonstrated potential in balancing resource consumption and monitoring efficiency lays a solid theoretical and practical foundation for building more efficient and intelligent blockchain P2P network monitoring systems in the future.