Error-Driven Prompt Optimization for Sustainable Tabular Reasoning

Abstract

The environmental cost of adapting Large Language Models (LLMs) is a growing concern in artificial intelligence research. We present a systematic comparison between gradient-based fine-tuning and a rule-based prompt optimization framework for financial tabular reasoning. Using the TAT-QA dataset, we demonstrate that our Error-Driven Code Generation Agent (CGA) achieves 66.45% accuracy on the Qwen3 4B model, surpassing GPT-3.5 Turbo. Crucially, this performance is attained with a fraction of the energy required for LoRA or full fine-tuning. Specifically, our proposed solution consumes, on average, >14x less electricity than fine-tuning approaches, significantly reducing operational overhead. Our results indicate that systematic prompt engineering offers a sustainable, privacy-compliant alternative for industrial AI deployment, achieving high precision without the carbon footprint of iterative training cycles. This work provides evidence for an energy-efficient (Green AI) alternative in regulated sectors, proving that strategic model adaptation can be both high-performing and energy-efficient.