Preprint reports adaptive LLM agent for chip-synthesis tuning

A preprint describes SynAct, a reasoning-and-acting LLM agent for logic-synthesis optimization. The system reads live synthesis reports, current circuit state, retrieved tool knowledge and prior optimization experience before issuing targeted commands. In experiments using a commercial synthesis tool across 14 designs, the authors report average worst negative slack fell to 27% of the bootstrap-synthesis result, while the agent balanced area and power trade-offs. The work has been posted as a preprint rather than independently verified research.