Supervised reward learning — is extended by → Semi-supervised or active reward learning

Supervised reward learning is the baseline instantiation of semi-supervised RL: train a model to predict reward from state, then use that prediction, weighted by confidence, to estimate payoff on episodes with no true label at all. The semi-supervised or active variant is explicitly framed as building on top of it — the paper introduces it with 'combine the above with traditional semi-supervised or active learning' (concrete-problems, §"5 Scalable Oversight", p. 12) — adding the ability for the agent to flag 'salient' events in the environment and request the scarce reward label specifically for those, rather than accepting whatever labels a passive random or fixed schedule happens to provide. The extension changes only how labels are allocated, not how they are used once obtained.