Energy-Aware Data-Driven Model Selection in LLM-Orchestrated AI Systems

Energy-Aware Data-Driven Model Selection in LLM-Orchestrated AI Systems Daria Smirnova, Hamid Nasiri, Marta Adamska, Zhengxin Yu, and Peter Garraghan School of Computing and Communications, Lancaster University Email: {d.smirnova1, h.nasiri, m.adamska, z.yu8, p.garraghan}@lancaster.ac.uk Abstract—As modern artificial intelligence (AI) systems become more advanced and capable, they can leverage a wide range of tools and models to perform complex tasks. Today, the task of orchestrating these models is often performed by Large Language Models (LLMs) that rely on qualitative descriptions of models for decision-making. However, the descriptions provided to these LLM-based orchestrators do not reflect true model capabilities and performance characteristics, leading to suboptimal model selection, reduced accuracy, and increased energy costs. In this paper, we conduct an empirical analysis of LLM-based orchestration limi- tations and propose GUIDE, a new energy-aware model selection framework that accounts for performance–energy trade-offs by incorporating quantitative model performance characteristics in decision-making. Experimental results demonstrate that GUIDE increases accuracy by 0.90%–11.92% across various evaluated tasks, and achieves up to 54% energy efficiency improvement, while reducing orchestrator model selection latency from 4.51 s to 7.2 ms. I. INTRODUCTION Modern AI systems have seen widespread adoption. Con- ventional AI systems are designed to complete a singular or narrow set of tasks such as object detection [1] or speech recognition [2]. AI systems – particularly those using Large Language Models (LLMs) – can facilitate complex tasks by leveraging an assortment of tools and models in tandem to expand system versatility and capability. Orchestration – responsible for determining model selection, planning, and execution to complete tasks [3] – is key in AI systems. As system scale and the number of models grow, there is an urgent need to design AI system orchestrators that operate at high speed, accuracy, and low cost. In response, researchers have recently identified that LLMs due to their strong reasoning capabilities can be instructed to operate as AI system orchestrators. These LLM-based orchestrators leverage foundational models to reason, make decisions, and act to invoke specific tools and models [4, 5, 6]. The effectiveness of these orchestrators has resulted in their increased adoption within agentic tool-augmented systems [7]. However, existing LLM-based orchestrators encounter many challenges pertaining to their performance and efficiency: (i) Orchestrator decision making for model selection is limited to qualitative data of a model’s capability, specifically model cards and textual descriptions (e.g., recommended tasks or number of user likes on HuggingFace) [8]. While such information helps provide task context to the orchestrator, it Tools Vision Model #1 LLM-based Orchestrator Planning Chain of Thought Reflection Memory Short-Term Long-Term Query Response Vision Model #2 Database Web Search Fig. 1. Overview of an LLM-orchestrated AI system. does not capture task-level performance (i.e. accuracy, energy cost). Reliance solely on qualitative model information results in suboptimal selection of models that are neither the most accurate nor cost-efficient. (ii) LLM-based orchestration requires extensive communi- cation with an LLM to operate — i.e., textual descriptions of models must be sent to the LLM to provide context, such as API descriptions and usage examples for each tool [9]. This incurs significant token usage and latency (especially at greater AI system scale), both of which increase energy consumption and reduce system throughput. Moreover, the LLM’s context window limits the number of tools the orchestrator can consider. Hence, we assert that the current design of LLM-based orchestrators for model selection significantly impairs AI system scalability, performance, and efficiency. To address these limitations, we postulate that incorporating quantitative data (accuracy, energy cost, etc.) into orchestrator decisions would result in more effective model selection, enabling the AI system to balance task accuracy and cost- efficiency (specifically energy use, given the growing concerns of AI sustainability and datacenter growth [10]). There have been extensive discussions on performance–energy trade-offs in adjacent domains such as LLM-serving [11, 12, 13, 14] and IoT [15]. However, to the best of our knowledge, there is no prior work that has analyzed the above limitations or the effectiveness of incorporating quantitative data within LLM- based orchestration for AI systems. In this paper, we present a comprehensive experimental analysis of LLM-based orchestrator operation and highlight recurrent shortcomings. We identify that LLM orchestrators arXiv:2512.01099v1 [cs.AI] 30 Nov 2025 systematically select underperforming models that increase costs and energy use with mini

This content is AI-processed based on open access ArXiv data.