AutoMCU: Feasibility-First MCU Neural Network Customization via LLM-based Multi-Agent Systems

2026-05-23 · Source: cs.LG updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Internet of Things (IoT) & Connected Devices, Emerging Technologies & Innovation · Depth: Expert, extended

Summary

AutoMCU is a novel feasibility-first large language model (LLM)-based multi-agent system designed for automated neural network customization on microcontroller units (MCUs). It addresses the challenge of deploying NNs on resource-constrained MCUs by iteratively generating architecture candidates, filtering infeasible designs via vendor toolchain feedback before training, and verifying deployability through backend-grounded analysis. The system incorporates hardware-in-the-loop architecture generation for early elimination of undeployable candidates under RAM and Flash constraints, and state-isolated multi-agent scheduling for stable coordination. Experiments on CIFAR-10 and CIFAR-100 under strict MCU constraints (e.g., RAM≤256KB, Flash≤512KB) demonstrate AutoMCU achieves competitive accuracy while reducing customization time to 1–2 hours, a significant improvement over hundreds of GPU hours for traditional MCU-oriented HW-NAS baselines like µNAS. Real-device deployments on STM32F407VET6 and STM32H723ZGT6 microcontrollers validate its practical applicability and adaptability. The system uses DeepSeek-V3.2 as its core LLM backend.

Key takeaway

For Machine Learning Engineers deploying neural networks on microcontrollers, AutoMCU offers a significantly faster and more reliable customization workflow. You can reduce model development time from hundreds of GPU hours to just 1–2 hours by leveraging its feasibility-first design. This approach ensures your models are not only accurate but also genuinely deployable on target MCUs, avoiding wasted effort on incompatible architectures. Consider integrating such a system to streamline your TinyML projects.

Key insights

AutoMCU automates MCU neural network customization by prioritizing backend-verified deployability before training.

Principles

• Feasibility is a hard constraint.
• Use structured architecture generation.
• Isolate agent states for stability.

Method

AutoMCU uses an LLM-based multi-agent system to propose, screen via vendor toolchains, train, and evaluate NN architectures in a closed loop, guided by historical feedback.

In practice

• Filter infeasible designs pre-training.
• Use vendor toolchains for validation.
• Employ structured LLM proposals.

Topics

• Microcontroller Units
• Neural Architecture Search
• Large Language Models
• Multi-Agent Systems
• Edge Intelligence
• TinyML

Code references

• XiaomiMiMo/MiMo-V2-Flash

Best for: AI Engineer, Computer Vision Engineer, Research Scientist, AI Scientist, Machine Learning Engineer, AI Hardware Engineer

Related on AIssential

• HPC Coding Agent - Part 3: MCP Tool for Profiling — An AI agent can autonomously profile and optimize GPU-accelerated HPC applications using a custom MCP server for tool integration.
• TI Adds NPU to a Third MCU Family — On-chip NPUs significantly enhance real-time control, predictive maintenance, and power efficiency in embedded systems.
• ChatNeuroSim: An LLM Agent Framework for Automated Compute-in-Memory Accelerator Deployment and Optimization — ChatNeuroSim automates CIM accelerator design and optimization using LLM agents and design space pruning for faster, more efficient exploration.
• Embedded Arena: Iterative Optimization via Hardware Feedback — Hardware-in-the-loop LLM agents can autonomously optimize embedded AI models, surpassing manual and feedback-free methods.
• MiCU: End-to-End Smart Home Command Understanding with Large Language Model — MiCU enhances smart home command understanding using LLMs, synthetic data, and specialized training for improved accuracy and efficiency.
• Agentic MPC for Semantic Control System Resynthesis — Agentic MPC combines LLM-based agents with modular MPC to enable context-aware, semantically adaptive control.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.LG updates on arXiv.org.