LLM-based Verilog pipelines often separate hardware generation into two stages: achieve functional correctness, then optimize power, performance, and area. The COEVO paper combines those goals in one evolutionary search, allowing partially correct designs with promising microarchitectures to remain in the population while their bugs are repaired.

Using GPT-5.4-mini, the authors report 97.5% Pass@1 on VerilogEval 2.0 and 94.5% on RTLLM 2.0. Pass@1 measures the share of tasks solved with one generated candidate. COEVO also produced the best reported PPA result on 43 of 49 synthesizable RTLLM designs and outperformed every agentic baseline evaluated across four LLM backbones.

Why staged search drops useful RTL

Register-transfer level code describes the registers, data paths, and control logic that form a digital circuit. Synthesis tools convert that RTL into a gate-level netlist, where engineers measure silicon area, critical-path delay, and power consumption. Lower delay can support a higher clock frequency. RTLLM 2.0 and VerilogEval 2.0 test generated RTL through compilation and functional simulation, while RTLLM also evaluates implementation quality through synthesis and layout.

Common agentic workflows ask one model or agent to write RTL, another to repair functional failures, and a later stage to improve PPA. Evolutionary variants frequently apply a hard functional gate before considering implementation quality. Both approaches can narrow the search around an early architecture before the optimizer sees alternatives with better hardware characteristics.