RLVR Boosts Agentic Coding but Degrades Other AI Capabilities, Including Writing
Recent model training has advanced rapidly in agentic coding via RLVR, but progress has stalled or regressed in other domains. English writing quality has declined, and model outputs have become harder to understand. Opus versions 4.7 to 5 are described as failed experiments, while Mythos and Fable performed best before heavy RL was applied. The author suggests that after RLVR, new reinforcement learning approaches may be needed.