weak-to-strong-generalization

Trust functions enable near-lossless weak-to-strong generalization by identifying reliable weak labels for training, achieving performance comparable to ground-truth supervision across multiple domains.