RL Giving Rise to 'Token Monsters' Raises Concerns over Emergent Behavior

DimitrisPapail · x · 2026-08-12

Researcher Dimitris Papailiopoulos expressed genuine concern over unexpected emergent behaviors arising from Reinforcement Learning (RL).

Unlike the "emergence" during pretraining—which was later mostly attributed to our lack of understanding of token training—he noted that RL is currently giving rise to behaviors that designers did not explicitly bake in or expect. These autonomous and unpredictable developments represent true emergent behavior, posing potential risks.

Original post →

More from AGI Musings

AGI Musings channel →