Conditional Feasibility Steering in Diffusion Models

Terminal verifiers can identify successful diffusion outputs but provide no guidance on unfinished states. We introduce Conditional Feasibility Steering (CFS), a fixed-budget discovery method that uses one completed pilot per saved state to allocate further generation. CFS favors states with successful pilots, independently reselects a state for each fresh continuation, and retains all completed outputs. A closed-form allocation rule bounds the second moment of the optional importance correction. No verifier gradient or learned value model is required. Our analysis connects the information in binary feedback to variation in conditional feasibility and separates the yield contributions of computation reuse and informed allocation. Controlled experiments in molecular and image generation show higher verified yield than uniform allocation within the same branching structure. Independent refresh reduces finite-target estimation error relative to single-parent commitment, although rejection remains more accurate. CFS is designed for verified discovery; its unweighted outputs are not claimed to follow the base conditional distribution.