Scalable oversight: PGR and the three mechanisms
Performance Gap Recovered (PGR)
weak accuracy ->
PGR
0.60
0.35
0.70
0.55
0.80
0.70
0.90
0.83
three scalable-oversight mechanisms
debate (Irving 2018)
two U's argue; weak judge picks. mixed 2024 evidence.
recursive reward modeling (Leike 2018)
U helps train RM for U+1. overseer tracks U.
task decomposition (Christiano 2018)
break hard task into checkable sub-tasks.
complementarity with W2SG
scalable oversight: improves label quality.
W2SG: closes gap from whatever labels.
combined (Lang 2025): better PGR on NLP tasks.
PGR is a measurable, not a solution. Burns et al. 2023 are explicit: the weak-strong gap they study is capability-shaped.
superalignment question -- can humans oversee superhuman models? -- depends on whether the gap transfers to alignment-shaped settings.
organizational: OpenAI Superalignment dissolved 05/2024; agenda continues at Anthropic, MATS, Redwood, Apollo, METR.