LLM co-evolution yields better approximate multipliers than EvoApproxLib
New algorithm uses off-the-shelf LLM to automate approximate multiplier design, beating benchmarks.
A new paper from researchers Martin Tomasovic and Lukas Sekanina presents a co-evolutionary algorithm that uses an off-the-shelf large language model (LLM) to automate the design of 8-bit approximate multipliers — circuits that deliberately relax accuracy to gain power, latency, and area efficiency. The approach simultaneously evolves two populations: candidate circuits and prompt templates that guide the LLM's modifications. Crucially, the LLM requires no domain-specific training, making the method accessible with current models. The algorithm targets error-resilient applications like neural networks, where small computational errors are acceptable.
Experimental results demonstrate that the co-evolutionary method discovers approximate multipliers with superior error-area trade-offs compared to the EvoApproxLib library, a well-known repository of optimized approximate circuits. The technique outperforms existing hand-tuned and evolutionary designs across multiple design objectives. This suggests that LLMs can serve as powerful tools for hardware design automation, reducing the manual effort needed to explore the vast design space of approximate computing. The paper is set to appear at the Parallel Problem Solving From Nature (PPSN) conference in Trento, Italy, in 2026.
- Co-evolves two populations: candidate circuit designs and LLM prompt templates.
- Uses off-the-shelf LLM without any domain-specific fine-tuning or training.
- Achieves 8-bit approximate multipliers with better error-area trade-offs than EvoApproxLib benchmarks.
Why It Matters
Automates hardware design for error-resilient AI chips, reducing manual optimization effort and enabling faster innovation.