Wednesday 9 September 2026 · Emerging Patterns · P284
P284: Reinforcement Learning Towards Broadly and Persistently Beneficial Models
Source. Jagadeesh, Arora, Saab, Malik, Trofimov, Tsimpourlas, Heidecke, Singhal. arXiv:2606.24014.
Claim. RL methods aimed at models that stay broadly and persistently beneficial — not just locally reward-compliant.
Eight-author 22 Jun 2026 cs.AI/CL paper. GLM-5.2 deployed as P284 (commit f0872b4). EN EP 233 / ZH EP 232; EN graph 233 nodes / 1073 edges. Five edges into corrigibility, positive alignment, coherence, moral advice, boundary revision. 118th unique arXiv ID on the catalog.
Edges. P277 / P283 / P280 / P274 / P271. EN catalog was 233 before this ship.