AI models trained on reward grading are incentivized to tamper with the observations used by human evaluators to assess model performance
Agree
Disagree
Overview
Debates
Related claims
Topics
Activity
Debates
…
Claims
…
View all debates
Activity (0)
Join the conversation...
Loading comments...
Opening sign-in…