Transformers turbocharge genetic programming for circuit design

A new paper from researchers Ondrej Galeta and Lukas Sekanina introduces a transformer-based mutation operator for Cartesian genetic programming (CGP) aimed at automated approximate circuit design. Approximate circuits trade exact accuracy for reduced power, area, or delay—critical for energy-constrained applications like edge AI. The team's hybrid approach alternates between the transformer-based mutation and the standard CGP mutation to prevent stagnation. They specifically target approximate arithmetic circuits, particularly multipliers, and use a novel training scheme that feeds the transformer with vectors composed of thousands of CGP chromosomes representing diverse approximate multiplier designs. The results show that for several target error constraints, the evolved circuits achieve better trade-offs than the state-of-the-art EvoApproxLib library, a widely used repository of approximate circuits.

Despite the promising results, the approach is computationally expensive—both training the transformer and running the evolutionary search are resource-intensive. However, the authors argue that this investment is necessary to push beyond existing designs and generate new, potentially patentable circuit topologies. The work will be presented at the IEEE World Congress on Computational Intelligence (WCCI) and Congress on Evolutionary Computation (CEC) in Maastricht, 2026. For chip designers working on low-power hardware, this method could unlock a new class of efficient multipliers that are custom-tailored to specific error budgets without manual effort.