Tag
This paper studies strong-to-weak capability transfer at test time, showing that stronger models can build inference-time harnesses that nearly double weaker models' performance without parameter updates.