weak-to-strong generalization

Weak-to-strong generalization asks a sharp question: if you only have a weak teacher whose labels are often wrong, can you still elicit the much larger latent competence of a strong student model? It is an empirical stand-in for the future situation where humans, who will be the weak supervisors, must align models that know more than they do. The hope is that the strong model already represents the right answer internally, so weak labels mostly need to point at that knowledge rather than teach it from scratch.

The canonical experiment fine-tunes a large pretrained model on labels generated by a much smaller model, then measures how much of the gap to ground-truth performance is recovered, summarized as a performance-gap-recovered ratio. Findings show strong students often outperform their weak teachers, generalizing past the teacher's mistakes, and that simple aids, an auxiliary confidence loss, bootstrapping through intermediate model sizes, and better unsupervised representations, can substantially raise recovery, though it remains far from full.

The result is encouraging but bounded, and it must be read with care. The strong model can also faithfully imitate the teacher's errors instead of correcting them, especially when the task rewards mimicking surface style over getting at the truth, and recovery turns out to be uneven across domains and prompt formats. So weak-to-strong is best read as early evidence that eliciting latent ability is possible at all, not as a finished alignment guarantee that scales to superhuman supervision.