weak-to-strong generalization
A weak-supervision research setting, and the generalization phenomenon it tests, in which a stronger learner is trained using imperfect labels or demonstrations from a weaker supervisor and is evaluated on whether it can outperform that supervisor, ideally approaching capabilities already available to the stronger learner; it is studied as an empirical analogue of humans supervising superhuman AI systems.
Metadata
- aliases
- en
- W2SG
- weak to strong generalization
- concept_id
- weak_to_strong_generalization
- preferred
- en
- weak-to-strong generalization
- relations
- relations
- broader
- prerequisite
- status
- active