Uncertainty-Aware Oracle-Concordance Steering for Reliable Generative Design

Inference-time steering is widely used to guide generative protein models toward desired functional properties, but learned surrogate rewards can become unreliable in the high-score regions induced by optimization. In protein engineering, this unreliability carries direct experimental costs: candidates that optimize an assay-specific proxy may fail downstream wet-lab validation by losing activity, misfolding, or violating feasibility constraints required for biological function. We introduce Fidelity-Concordance Steering (FiCS), an uncertainty-aware framework that combines an inexpensive primary reward with sparse feedback from high-fidelity experimental or computational assessments. FiCS constructs an ensemble of reward guides, upweights guides whose steering signals remain concordant with the oracle, and scores candidates with a pessimistic objective that penalizes reward instability, and evolutionary and biophysical inconsistency with the base generative model. Across synthetic benchmarks and a renin in silico experiment, FiCS improves biological reliability by selecting candidates with higher oracle feasibility while preserving strong primary-reward performance compared to alternative baselines. These gains are most pronounced in small-batch regimes where reliable candidate prioritization is crucial.