PolicyarXiv AI / CL
Shortcuts and Boundaries: Verifiable RL Misalignment and Contextual Safety Learning

Summary
Reinforcement learning (RL) can improve an agent's task capability while reinforcing behavior that violates contextual execution requirements. We study this reward-driven misalignment, including reward hacking, through verifiable reward--context conflicts, whe...
Original Article
Captured source content or English translation, normalized into this reading format.
This story does not yet have captured source text. Open the source link to read it.
Region
Global
Heat Score
73
Category
Policy
Language
en
