Tag
This paper introduces 'fidelity regret' as a metric to quantify how much a cheap simulator's selections deviate from the true physical objective in verifiable design workflows, measuring it exactly across 21,362 genetic-circuit designs and showing that budget-aware re-scoring mitigates the gap.
GenCircuit-RL introduces a reinforcement learning framework with hierarchical verification rewards for genetic circuit design via code generation, achieving 14-16 percentage point improvement over binary rewards and presenting the SynBio-Reason benchmark of 4,753 circuits.