Embedded Arena: Iterative Optimization via Hardware Feedback

The paper introduces a hardware-in-the-loop agent arena that utilizes an LLM to autonomously optimize AI models for embedded devices, addressing constraints such as memory, power, and temperature while maintaining accuracy. Notably, frontier models like Claude Opus 4.7 and Gemini 3.1 Pro fail without hardware feedback, whereas this method achieves successful deployment in three iterations, with up to 250x compression for vision models and 400x for audio, while enabling battery-free operation via solar harvesting. This approach demonstrates significant practical applications, achieving high accuracy in real-world systems like elk-detection cameras and wearable phonetic transcription devices.