arXiv: MARLIN — Multi-Agent Game-Theoretic RL Cuts LLM Inference Carbon 33%, Water 43%, Energy 11%

arXiv·medium signal

Paper 2605.13496 (May 13) presents MARLIN, a game-theoretic reinforcement learning framework for sustainable LLM inference in cloud datacenters that co-optimizes TTFT, carbon emissions, water usage, and energy costs. Results: 18% TTFT reduction, 33% carbon reduction, 43% water reduction, 11% energy cost reduction vs. state-of-the-art. Key finding: LLM inference now accounts for 90% of total LLM lifecycle energy, dwarfing training costs. MARLIN operates as a meta-scheduler above Kubernetes/vLLM containers.