Logic synthesis tuning is high-dimensional and expensive: PPA results are highly sensitive to the sequence of optimization commands, the search space is enormous, and the only feedback is the synthesis report after each run. Previous automated approaches either black-box search a fixed action space (no interpretability) or use an LLM to generate a static script upfront (no adaptation). SynAct closes the loop. It reads the live synthesis report after each step, diagnoses the current circuit state against retrieved tool knowledge and historical optimization experience, and issues a targeted command for the next step. The result is worst negative slack reduced to 27% of what bootstrap synthesis produces, across 14 designs on a commercial synthesis tool.
The mechanism is the diagnosis-and-targeting loop, not the LLM. A static script cannot respond to a timing closure that half-worked and left specific paths critical. A search over a fixed action space cannot reason about why a specific path is violating or which optimization addresses it. SynAct reads the report, identifies the critical paths and their root causes, and selects the command most likely to fix that specific state. That is the workflow a senior implementation engineer runs manually, now running autonomously and iterating until WNS converges.
The near-term application is implementation runs that currently require an expert to babysit closure. A team running 50 design experiments simultaneously cannot staff an expert per run. An agent that diagnoses and targets autonomously can. The longer-term implication is that timing closure is no longer a gate that serializes the design iteration loop -- it becomes a background process that runs to completion while the team moves to the next experiment.