SKILL: Self-correcting Knowledge-guided Iterative Large Language Model Agent for Logic Optimization
06:00 · August 18, 2026 · arXiv cs.AI RSS
Logic synthesis optimization poses significant challenges due to exponentially growing search spaces, sparse reward signals, and diverse logic structures. Traditional expert-designed flows lack adaptability, while reinforcement learning (RL) methods often suffer from low sample efficiency and limited interpretability. We introduce SKILL, a Self-correcting Knowledge-guided Iterative Large Language Model Agent that unifies multi-agent LLM reasoning and RL-based environment interaction for automated synthesis optimization. SKILL coordinates three specialized LLMs: GPT-4o for strategic planning, Claude Sonnet 4 for detailed reasoning, and Gemini 2.5 Pro for efficient analysis with a PPO-based RL agent that learns actionable policies through direct interaction with synthesis tools. A novel self-correcting module monitors environment feedback (PDA metrics), detects suboptimal behaviors, and invokes LLM-guided recovery strategies. Evaluations on IWLS, OpenCores, and EPFL benchmarks show SKILL achieves a 12.4 % PDA improvement over expert flows and 86.3% success rate on logic systems up to 500K gates.
Summary
Logic synthesis optimization involves navigating vast combinatorial spaces of logic transformations to improve circuit metrics such as power, delay, and area. Conventional expert-designed scripts in tools like ABC and Yosys deliver solid results on familiar designs but adapt poorly to new topologies or technology nodes. Reinforcement-learning approaches address some of this rigidity by learning policies through direct interaction with synthesis environments, yet they often encounter sparse rewards and produce policies that are difficult to interpret or debug.
SKILL integrates three large language models with a proximal-policy-optimization agent to combine high-level reasoning and low-level tool control. GPT-4o generates strategic plans, Claude Sonnet 4 performs detailed step-wise analysis, and Gemini 2.5 Pro supplies fast structural checks. The PPO agent translates these directives into concrete synthesis operations and receives immediate Power-Delay-Area feedback. A dedicated self-correction loop continuously compares observed PDA values against expected progress; when regressions appear, it triggers the language-model ensemble to diagnose the failure and propose recovery actions.
The resulting closed-loop system was evaluated on IWLS, OpenCores, and EPFL benchmarks containing designs up to 500 000 gates. Across these suites SKILL recorded an average 12.4 % PDA reduction relative to established expert flows while maintaining an 86.3 % success rate. The architecture therefore demonstrates that coordinated language-model guidance and reinforcement-learning interaction can improve both solution quality and robustness in industrial-scale logic synthesis.
Why it matters
This research is highly relevant for the Dutch AI and semiconductor ecosystem, offering advanced AI methodologies to optimize chip design and logic synthesis. It provides researchers with a rigorous, novel approach combining multi-agent LLMs and RL that can be directly applied to industrial-scale electronic design automation.
