Abstract Activation Spaces for Content-Invariant Reasoning in Large Language Models | ScienceToStartup | ScienceToStartup