Tag
A reproducibility study of Khatri et al.'s latent-space safety probes, testing generalization across model families and sensitivity to non-determinism. Results show the probes extend to other models with similar F1 scores, and final token latent vectors remain consistent across seeds.