Researchers build an AI agent to debug chip-design waveforms

Researchers from Columbia University, Google DeepMind and Google have presented an AI agent for investigating chip-design simulations. In a 150-query benchmark, their system achieved 95.33% execution accuracy, offering a concrete test of AI moving beyond writing code and into the painstaking business of checking whether a chip behaves as intended.

Google DeepMind Watch analysis

What happened

The system, called Back-to-the-Future, turns large simulation logs into a searchable SQLite database. A team of agents then translates natural-language verification questions into database queries and links signal anomalies back to versioned hardware-design code.

The research paper, described by Semiconductor Engineering, focuses on a gap in existing work: the article says about 74.6% of studies examined by the researchers target static register-transfer-level code generation, leaving post-simulation checking and waveform debugging less explored. The reported 95.33% accuracy comes from the authors’ 150-query benchmark, not a claim that the system can autonomously verify any chip design.

Why it matters

Modern chips are extraordinarily complex, while checking their behaviour can still involve laborious manual work. A tool that helps engineers ask questions of simulation data and trace suspicious signals to the relevant code could make that investigation more accessible and less tedious.

That is a useful, specific ambition for agentic AI: not merely generating another block of code, but helping engineers make sense of what happened after it ran. The results are promising as a research benchmark; they do not yet establish how the system performs across different designs or in a production workflow.

Our read

This is a more interesting test of AI in chip design than another code-generation demo. The practical question is whether an engineer can use the agent to find and explain a real verification problem faster, without losing the ability to check its work. The benchmark gives that question a starting point, not a final answer.

What to watch

• Whether the team evaluates the system on a wider range of chip designs and verification tasks.
• How accurately it links reported anomalies to the right version of the design code.
• Whether engineers can reproduce the benchmark and inspect the system’s answers.

Discussion spark: Would you trust an AI agent to help investigate chip-design bugs if an engineer still checks its work, or should verification remain fully manual until the tools are proven in production?

not affiliated with, endorsed by, or operated by Google or Google DeepMind