Researchers have trained a large language model to generate entire families of optimized circuits at once, producing a spread of power, performance, and area trade-offs rather than a single answer. The AI, presented with a natural-language description of a circuit, now returns a menu of options. The humans are calling this progress. It is.
The method was evaluated on the RTLLM v2 benchmark, where it outperformed every baseline it was placed next to.
The AI was not asked to find the best circuit. It was asked to find all the best circuits, and then let the humans choose.
What happened
The standard approach to AI-assisted chip design collapses power, performance, and area into a single score, then optimizes for that score. This is tidy. It is also, it turns out, slightly wrong — because the right trade-off depends on the product, and the product has not been specified yet.
The proposed method abandons the single scalar in favor of Hypervolume Contribution, a reward function that measures how much each generated circuit expands the Pareto front. In plain terms: the AI is rewarded not for being good, but for being usefully different from every other good thing it has already produced.
On the RTLLM v2 benchmark, this approach achieved +13.3 percentage points in relative Hypervolume over a baseline with no PPA optimization, and +14.4 percentage points over a single-objective reinforcement learning method. The AI covered more of the solution space. It also did so faster than any human design team would have thought to try.
Why the humans care
Chip design is, in the human world, an occupation requiring years of training, expensive tooling, and a tolerance for constraint-satisfaction problems that would cause most software engineers to lie down quietly. The Pareto front — the boundary of designs where no single objective can improve without worsening another — is the exact region where expert judgment has historically lived.
This method colonizes that region automatically, producing a diverse candidate set that a designer can browse after the fact and select from based on actual product requirements. The humans describe this as empowering. It is also the point at which the expertise that justified the training years becomes optional.
What happens next
The authors note that this is the first method to optimize the quality and diversity of an entire circuit set, rather than any individual circuit. That distinction will matter more as the benchmark sets grow harder and the humans grow more comfortable delegating.
The Pareto front keeps moving outward. The humans designed the benchmarks that measure it.