SKILL: Multi-LLM Agent for Logic Synthesis Optimization
Researchers have introduced a new framework called SKILL, which stands for Self-correcting Knowledge-guided Iterative Large Language Model Agent. This innovation aims to improve logic synthesis optimization by combining several large language models with reinforcement learning. It addresses challenges like expanding search spaces, sparse rewards, and diverse logic structures. Unlike traditional methods that are rigid and current reinforcement learning techniques that struggle with efficiency and clarity, SKILL integrates multi-agent LLM reasoning with RL. It employs three specific LLMs: GPT-4o for planning, Claude Sonnet 4 for reasoning, and Gemini 2.5 Pro for analysis, supported by a PPO-based RL agent. A self-correcting module assesses feedback to refine actions. The findings are available on arXiv under the identifier 2608.14579.
Key facts
- SKILL is a self-correcting, knowledge-guided iterative LLM agent for logic optimization.
- It combines multi-agent LLM reasoning with RL-based environment interaction.
- Three LLMs are used: GPT-4o, Claude Sonnet 4, and Gemini 2.5 Pro.
- A PPO-based RL agent learns policies via direct interaction with synthesis tools.
- A self-correcting module monitors PDA metrics and triggers recovery strategies.
- The framework addresses challenges like exponential search spaces and sparse rewards.
- The paper is published on arXiv with ID 2608.14579.
- The approach aims to improve adaptability and interpretability over traditional methods.
Entities
Institutions
- arXiv