Causal influence in latent spaces depends on capacity, responsiveness, and alignment

Causal effectiveness of latent structures in language models is determined by three distinct constraints: capacity (output sensitivity), responsiveness (contextual promotability), and alignment (context-specific representation). Empirical tests across 50 concepts show that low capacity or responsiveness reduces causal influence by 84% and 95% respectively.