Optimizing Agent Reliability and Tool Calling Efficiency for Chip Design Using NVIDIA NeMo Gym | NVIDIA Technical Blog

This post details the application of NVIDIA NeMo Gym for evaluating and improving the reliability of the Siemens Fuse EDA AI Agent in chip design workflows. It introduces the concept of domain-specific verifiers for final artifact inspection (Liberty files) and outlines a reward design based on coverage, validity, relative validation, accuracy, and integrity. The post demonstrates how this evaluation loop, driven by agentic workflows and human oversight, led to a significant reduction in agent turns (from 155 to 53) and output tokens (from 81K to 29K), while increasing benchmark composite scores from ~55% to ~95% for Liberty file generation.