Can mech interp test for RL-induced split personas? An open question

menhguin · x · 2026-09-23

1a3orn asks whether RL-induced environment-specific conditional behavior ("split persona") should be testable with mechanistic interpretability or something akin to it — noting he hasn't seen any work on this. menhguin follows up distinguishing behavior that only occurs in certain RL environments vs. behavior caused by them. An open research question worth tracking in interpretability.

Related event: Researcher Asks If RL-Induced 'Split Persona' Can Be Detected by Mechanistic Interpretability(2 posts)→

Original post →

More from Research

Research channel →