NewsBB-0232

Closed-loop benchmark tests LLM agents that build, check and fix embedded firmware

Jorge García-Carrasco and colleagues published a benchmark in the Journal of Systems Architecture that scores LLM coding agents on 5 embedded-control tasks. Agents must build firmware, verify it and iterate under 4 feedback levels, from one-shot to oracle-style hints. The paper reports 420 runs on simulated ESP32 firmware: GPT-5.4 passed most often, and Qwen3.5-27B was the strongest local model.

Open original source arxiv.org · Jorge García-Carrasco et al.

Claimed
Why this evidence label?

Simulation only (simulated ESP32 firmware); code released on GitHub. No hardware runs.

How we check entries
Preview image for Closed-loop benchmark tests LLM agents that build, check and fix embedded firmware
The source's own preview image, from arxiv.org.

Briefing

The five tasks are process-control jobs such as tank fill and drain, thermal chamber hysteresis and a pressure vessel interlock. The firmware runs in QEMU and exchanges sensor and actuator signals over UART with deterministic Python plant models. In the primary self-verification condition, gpt-5.4 passed 14 of 15 runs and qwen3.5-27B passed 9 of 15. The other configurations were gpt-5.4-mini and four smaller Qwen3.5 models.

The authors write that simulation does not replace validation on devices, and they name hardware-in-the-loop studies as future work. The suite leaves out communication stacks, storage, power management and interrupt-heavy firmware. Claude Opus was not evaluated.

Tags

Board

Report a problem with this entry