CERBERUS: An Autonomous, Self-Evolving LLM Agent System for Design-Adaptive QoR Optimization

作者：Tao Zhou, Jie Zhou, Jie Zhou · 发表于：Zenodo (CERN European Organization for Nuclear Research) · 年份：2026 · DOI：10.5281/zenodo.21623209 · 研究领域：Computer science、Real-time computing、Programming language、Algorithm、Embedded system、Arithmetic、Simulation

CERBERUS is the three-agent autonomous optimization system described in the paper: an optimizer that performs the iteration, a disruptor that is woken on a plateau in a fresh session and must return a plan, and a coder that is the only agent permitted to modify OpenROAD source. The agents run on Claude Code and communicate only through files, so the loop survives context resets and is bounded by wall-clock time rather than by a context window. The artifact contains the complete system: the four skills that drive the loop, the two sub-agent role prompts, the prior knowledge given to the agents, the OpenROAD baseline/evaluation/scoring flows, and a logic-equivalence gate that invalidates any iteration whose netlist no longer computes the same function. Because the prompts are the system, they ship as source rather than as a description. The mandatory evaluation invokes no language model. What the loop produced is deterministic — eight per-design optimizers, frozen exactly as submitted to the contest — and ae/run.sh smoke runs one of them end to end inside the contest Docker image, scores it with the official evaluator, and compares against the recorded result. On aes_cipher_top this takes about six minutes and reproduces the quality score exactly (Sfinal_noR 58.5021, measured twice from a clean container), with networking disabled and without a GPU, a commercial tool, or an API key. ae/run.sh verify re-scores an existing output in about twelve seconds, and ae/run.sh full extend...