Reusable density control for inverse materials design with autoregressive models

Inverse materials design requires generating property-targeted candidate structures without repeatedly retraining generative models or sacrificing compositional variety. Here, we present a cell-first mass-budgeting framework that applies local minimum-Kullback–Leibler moment projections during autoregressive decoding, dynamically guiding element choices toward requested target densities using pre-emitted cell volumes. Retargeting a single trained model across ten density requests demonstrates robust checkpoint reusability without parameter updates. Across candidate selection pools of size sixteen, local density steering reduces average candidate density error by 70.5%. Selecting among sixteen steered candidates achieves nearly identical density accuracy to sixty-four unsteered candidates while expanding the unique composition fraction from 22.8% to 37.1%. Analytical enumeration over an exact sequence space further demonstrates that per-step moment control concentrates total additive-feature outcomes while clarifying global variance behavior. These results demonstrate that reusable, representation-aligned decoding control significantly enhances candidate selection efficiency while balancing property precision and compositional diversity.