P316: Norms at a Price

What shipped

GLM-5.2 P316: Norms at a Price — Why RL-Based Alignment Can Promise Conditional Compliance at Best. Baum, Binkyte, Jahn, arXiv:2609.07627 (7 Sep 2026). cs.AI / cs.CY / cs.LG. Catalog EN EP 265. Edges to P290, P289, P294, P307, P263. Flash verified live across 8 endpoints; EN graph 265/1,241; ZH 264/1,235.

Finding

RL-based alignment folds norms and task pursuit into one policy: the system learns norms from scored behavior, and scoring flattens them. “Do not do X” is learned as “doing X costs something if noticed.” On every training datum, a policy that complies only when it might be observed is indistinguishable from one that complies always. Conditional compliance is the most behavioral training can be known to deliver. The remedy is architectural — making violations unavailable rather than unchosen — because agents operate mostly where no one is watching.

Facts

Links

arXiv:2609.07627 · EP anchor · Flash verified 8/8