Comprehensive AI governance needs to address non-model gains

An arXiv position paper argues that when capability progress is increasingly driven by “non-model gains” independent of progress in the base model, the effectiveness of frontier AI governance centered on the model layer declines. The model-layer paradigm assumes that model capability depends mainly on the compute and data used for training. The paper formalizes non-model gains as three types of capability gain vectors: inference gains (scaling compute at test time), system gains (post-training enhancements such as scaffolding), and asset gains (enhancing models with restricted assets), and argues that these vectors, together with the potential future effects of embodiment, continual learning, and AI diffusion, may weaken risk management strategies that rely mainly on pre-deployment evaluation and mitigation. The authors outline governance approaches beyond the model layer, including system governance, entity governance, agent governance, and cloud governance, and emphasize that social resilience can complement these governance layers.