Alignment Instability Condition

A three-part geometric condition — low-rank sensitivity, initial orthogonality, and curvature coupling — introduced by Springer et al. (2026) to explain why benign fine-tuning degrades LLM safety even with no harmful training data: early updates avoid the alignment-sensitive subspace, but loss-landscape curvature systematically steers later trajectories into it. Stubbed at watching (per the entity spec's emerging-term rule: first_seen cannot be backfilled) because it is a coinage from a single unrefereed 2026 preprint; graduates to hub only if the curvature-drift mechanism recurs in independent work.

References

• claim-springer-2026-finetuning-orthogonality-unstable-curvature-steers-into-alignment-subspace