SKILL Agent Optimizes Logic Synthesis with Self-Correction and LLMs

Logic synthesis optimization, a crucial step in chip design, faces significant challenges due to vast search spaces and complex reward signals. Traditional methods often lack adaptability, while reinforcement learning (RL) approaches can be inefficient and difficult to interpret. A new framework, SKILL (Self-correcting Knowledge-guided Iterative Large Language Model Agent), has been developed to address these issues. SKILL integrates multiple specialized Large Language Models (LLMs) for strategic planning, detailed reasoning, and efficient analysis, alongside an RL agent that learns through direct interaction with synthesis tools. A key innovation is its self-correcting module, which monitors environmental feedback and triggers LLM-guided recovery strategies when suboptimal behaviors are detected. Evaluations on standard benchmarks demonstrated SKILL's effectiveness, achieving a 12.4% improvement in performance over expert-designed flows and an 86.3% success rate on logic systems up to 500,000 gates.