Toward Sustainable Agentic AI: Carbon Tradeoffs Across Iterative Self-Refinement and Tool-Augmented Coding
A new arXiv paper presents two empirical studies estimating the energy and carbon footprint of agentic language-model systems. In iterative self-refinement on MATH, three refinement steps closed a 16-percentage-point acc...