ORACLE: A Multi-Objective Reinforcement Learning-Based Analog Circuit Design Optimizer with Large Language Models-Guided Exploration
ORACLE uses preference-conditioned vector rewards so one trained optimizer can handle different analog-circuit tradeoffs without retraining.
The paper says existing RL methods often collapse multi-objective specs into one scalar reward, losing Pareto tradeoffs. ORACLE adds normalized-weight and cosine-aligned preference guidance, plus LLM-guided action filtering to avoid poor or slow design moves. Across multiple circuit topologies and 2,000 test cases, the authors report 20.4x to 104.4x runtime reductions, 99.9% target-spec compliance, and 5.1x to 318.6x better figure of merit. Source: ArXiv · AI/CL/LG's note
The paper says existing RL methods often collapse multi-objective specs into one scalar reward, losing Pareto tradeoffs. ORACLE adds normalized-weight and cosine-aligned preference guidance, plus LLM-guided action filtering to avoid poor or slow design moves. Across multiple circuit topologies and 2,000 test cases, the authors report 20.4x to 104.4x runtime reductions, 99.9% target-spec compliance, and 5.1x to 318.6x better figure of merit. Source: ArXiv · AI/CL/LG's note
score 5