Beyond the Need for Speed: Energy-Aware Code Generation via Simulation-Guided Reinforcement Learning

2026-07-05

AI Summary

This work addresses the neglect of energy efficiency in existing code generation models and the impracticality of large-scale, reproducible hardware-based energy feedback. To bridge this gap, the authors propose the first simulation-based framework for energy-efficient code generation, featuring Green Tea—a deterministic architectural simulator enabling 3.5 million energy evaluations of C++ code snippets. They train energy-aware models via supervised fine-tuning followed by closed-loop reinforcement learning with a novel algorithm (GRPO), and introduce the CARET metric to jointly assess functional correctness and energy efficiency. Experiments on 143 held-out problems demonstrate a 12.63% CARET improvement over baselines, with generated code outperforming human expert implementations in energy efficiency by 58.4%. The study also reveals the misleading nature of traditional throughput-oriented metrics like IPC for energy ranking. The open-sourced dataset and infrastructure eliminate approximately 263,000 CPU hours of reproduction costs.

Abstract

Code models strictly prioritize functional correctness, leaving software energy efficiency as an unoptimized byproduct. Training models to generate energy-efficient code requires reproducible feedback at scale, which physical hardware measurement cannot reliably provide due to variance. In this paper, we replace hardware profiling with a deterministic architectural simulation harness to build Green Tea, a corpus of $3.5$ million evaluations across $1{,}474$ C++ problems. We train an energy-aware code model via supervised fine-tuning on energy-contrastive pairs, followed by closed-loop reinforcement learning (GRPO) using simulation-in-the-loop feedback. To rigorously evaluate deployment readiness, we introduce the Correctness-Adjusted Reduction in Energy Total (CARET), a metric that explicitly penalizes code that sacrifices functionality for efficiency. On $143$ held-out problems, our simulation-in-the-loop pipeline achieves $12.63\%$ CARET, nearly tripling the gain of fine-tuning alone, and successfully beats the energy efficiency of human-expert references on $58.4\%$ of its valid outputs. Furthermore, our analysis exposes the IPC trap: standard throughput proxies like Instructions-Per-Cycle (IPC) actively misrank true energy efficiency on $67.8\%$ of problems, proving the absolute necessity of direct energy simulation. By releasing our dataset and infrastructure, we bypass the $263{,}000$ CPU-hours required for reproduction, structurally empowering the community to deploy inherently energy-efficient code generation models.

Problem

• energy-efficient code generation
• code model
• hardware measurement variance
• energy simulation
• functional correctness

Innovation

• energy-aware code generation
• simulation-guided reinforcement learning
• architectural simulation
• CARET metric
• Green Tea dataset