Announcement
Harbor v0.25.0 introduces features aimed at mitigating reward hacking
A commentator says tools should help evaluation engineers develop robust tasks and verifiers as organizations improve at defining benchmarks.
TLDR
Harbor announced new features in v0.25.0 intended to mitigate reward hacking. A commentator argued that as organizations get better at defining benchmark tasks, developing robust tasks and verifiers will become the more challenging problem, and tools should help evaluation engineers do that work.
Combined views
711
4 Sources, first seen ago
