Enabling Cloud-Level Accuracy in Edge AI through IoT Data Preprocessing

Research Approach

This paper addresses the challenge of deploying Large Language Models (LLMs) for IoT environmental monitoring. While cloud-based LLMs offer high reasoning capabilities, they introduce latency, privacy, and connectivity risks. Conversely, compact edge-based models often struggle with raw numerical sensor data. The authors propose a structured prompt-side preprocessing framework that converts raw sensor readings into increasingly enriched textual formats: raw values, threshold-aware descriptions, and status-enriched summary flags. By shifting the burden of numerical interpretation from the LLM's internal reasoning to the prompt construction layer, the authors aim to improve the accuracy-latency trade-off for edge AI.

Experimental Setup

The researchers evaluated five local and five cloud LLMs using indoor datasets (Raspberry Pi/BME680 sensors) and outdoor air-quality datasets from multiple European cities. They constructed a binary query dataset covering air quality, thermal comfort, and joint environmental conditions. The models were tested under two inference modes—No-Chain-of-Thought (No-CoT) and Chain-of-Thought (CoT)—across three levels of prompt enrichment to measure both classification accuracy and response latency.

Key Findings

The results demonstrate that prompt enrichment is a highly effective strategy for local LLMs. In the No-CoT configuration, accuracy for local models improved from 50.9% to 81.7% for indoor tasks and from 63.7% to 89.3% for outdoor tasks when moving from raw data to the most enriched prompt format. Furthermore, the study highlights that the No-CoT mode with enriched prompts provides the most efficient configuration, achieving a mean latency of approximately 0.22 seconds, whereas CoT significantly increases inference time without necessarily providing proportional gains in this specific classification context.

This research provides a practical pathway for deploying intelligent, natural-language-based environmental monitoring at the edge. By demonstrating that simple, structured preprocessing can bridge the performance gap between compact local models and powerful cloud models, the study offers a scalable solution for privacy-sensitive and latency-critical smart environments, such as smart buildings and industrial monitoring systems.