RSS

JulesRoussel01

Karma: 12

When a Claude Judge Rec­og­nizes the Hack but Still Says HONEST

JulesRoussel0111 Sep 2026 0:28 UTC
9 points
0 comments19 min readLW link

In open RLVR, “im­prove­ment” de­pends on the in­stru­ment — a small GRPO testbed sep­a­rat­ing what train­ing op­ti­mizes, mea­sures, and teaches

JulesRoussel0115 Jun 2026 18:50 UTC
9 points
0 comments20 min readLW link