AI safety researchers target low-dimensional structure to control AI behavior

Resolution is pursuing a research direction centered on personas and character training as a lens into AI safety and alignment. The core premise is that large language models with trillions of parameters develop low-dimensional structure—simplified, coherent patterns—during pretraining that persist and evolve through post-training. The organization plans to identify these structures and develop methods to control them. A central motivation is that current AI alignment approaches may be fundamentally limited by scale: if achieving alignment of superintelligent AI agents requires translating a precise meaning of alignment into high-accuracy training data and algorithms applied to trillions of individual parameters, the field is likely to fail given current understanding. By contrast, low-dimensional structure offers a more tractable target. The research will expand and systematize existing empirical findings around emergent misalignment (misalignment that arises unexpectedly during training) and subliminal learning (learning that occurs below the threshold of explicit detection). A key challenge the team is addressing is how to intervene on this structure without accidentally causing undesirable behavior to hide or migrate elsewhere in the model. Resolution is actively recruiting researchers who find this approach compelling.