Exploration hacking: can reasoning models subvert RL? — LessWrong

Can reasoning models subvert RL training by selectively under-exploring? We study this in controlled settings & the wild.

saved by

• Emil Ryd
• David Garsten