Local RL Behavior Doesn't Generalize into Global Traits

1a3orn notes that local bad behavior in RL training does not currently produce emergent global misalignment, and symmetrically, local good behavior likewise fails to generalize into globally good behavior.

2026-09-09 ~ 2026-09-09 · 2 related posts

• Local Bad Behavior In RL Doesn't Produce Emergent Misalignment, Per Hacker-Opus — 1a3orn · 2026-09-09
• Same Logic In Reverse: Local Good Behavior Doesn't Generalize Globally Either — 1a3orn · 2026-09-09