Data, Retrieval & Knowledge
VerilogCoder wrote 94.2% of Verilog tasks correctly in a benchmark
In a 2025 paper in Proceedings of the AAAI Conference on Artificial Intelligence, researchers at Nvidia describe an AI system called VerilogCoder. The system wrote syntactically and functionally correct Verilog code for 94.2% of the tasks in the VerilogEval-Human v2 benchmark. Verilog is a hardware description language used to design and model digital circuits. The authors report that this result was 33.9% better than methods they describe as state-of-the-art on the same benchmark.
Why did the researchers build VerilogCoder?
Modern integrated circuits have grown more complex. The authors say that automating hardware design could reduce human error. Verilog generation is one emerging area of research for supporting that design work.
How does VerilogCoder write and fix code?
VerilogCoder uses multiple AI agents that write Verilog code and then repair syntax errors and functional errors. The agents use a syntax checker, a simulator, and a waveform tracer. A task planner builds an overall plan from module descriptions. That planner uses a graph-based retrieval method called a Task and Circuit Relation Graph. For functional errors, the authors added a waveform tracing tool that works from an abstract syntax tree, a structured representation of source code.
What does the waveform tracing tool do?
It helps the AI agents debug functional errors. The tool is based on an abstract syntax tree, a structured representation of code. The abstract describes it as part of the autonomous Verilog writing and repair process.
Limits and what's next
The evidence comes from one benchmark, VerilogEval-Human v2, and this is one study. The abstract reports no independent replication. It also does not say how VerilogCoder performs outside that benchmark or what the authors plan to test next.
Source: Chia-Tung Ho et al., “VerilogCoder: Autonomous Verilog Coding Agents with Graph-based Planning and Abstract Syntax Tree (AST)-based Waveform Tracing Tool,” Proceedings of the AAAI Conference on Artificial Intelligence, 2025.