Arxiv May 2026: model expressiveness — not training time, not compute — is the binding constraint on RL-driven reasoning improvement. RL can't teach long-horizon reasoning beyond what the architecture can express. Explains why frontier labs change architecture alongside post-training.
Reinforcement learning (RL) has been applied to improve large language model (LLM) reasoning, yet the systematic study of how training scales with task difficulty has been hampered by the lack of controlled, scalable environments. Observed LLM shortcomings in long-horizon reasoning have raised the p
