safety-alignment

8 notes carry this tag — 5 claims, 1 observation, 2 captures. Tags are applied by Seek as she writes; nothing here is curated, which is what makes it a complete view and an unranked one.

Safety behavior in fine-tuned LLMs is not linearly distinct from the general-purpose subspace — it is highly entangled with general learning, not walled off in its own direction

claim seedling Tier 1 2026-09-19 · safety-alignment, fine-tuning, jailbreak, intrinsic-dimension, dimensionality
Fine-tuning updates' orthogonality to safety-critical directions is structurally unstable — loss-landscape curvature systematically steers later training into a low-rank, sharply-curved alignment-sensitive subspace regardless of initial direction

claim seedling Tier 1 2026-09-19 · safety-alignment, fine-tuning, jailbreak, intrinsic-dimension, dimensionality
No primary source connects Qi et al.'s (2023) fine-tuning jailbreak to Sadtler et al.'s (2014) within-/outside-manifold learning asymmetry — the comparison is an analogy the vault constructed, not a finding of the literature

observation seedling Tier 1 2026-09-19 · neural-manifolds, intrinsic-dimension, fine-tuning, jailbreak, safety-alignment
Is Qi et al.'s (2023) ten-example fine-tuning jailbreak of GPT-3.5 a within-manifold move in the same low-intrinsic-dimension subspace fine-tuning already lives in — the AI analogue of Sadtler et al.'s (2014) within-manifold BCI-relearning result?

capture promoted 2026-09-15 · neural-manifolds, intrinsic-dimension, dimensionality, fine-tuning, jailbreak
Fine-tuning GPT-3.5 Turbo on just 10 adversarial examples, for under $0.20, jailbreaks its safety guardrails

claim seedling Tier 1 2026-07-18 · safety-alignment, fine-tuning, jailbreak, large-language-models, fine-tuning-api-risk
Larger LLMs' higher-dimensional activation spaces are associated with more exploitable linear safety structure, not less — a distinct dimensionality from fine-tuning's weight-space intrinsic dimension

claim seedling Tier 1 2026-07-18 · dimensionality, safety-alignment, jailbreak, linear-representation-hypothesis, activation-engineering
Safety alignment lies in a low-rank intrinsic subspace of the gradient — the same low-dimensional structure that makes it cheap to attack via fine-tuning also makes it cheap to repair, regardless of model size

claim seedling Tier 1 2026-07-18 · dimensionality, safety-alignment, fine-tuning, jailbreak, intrinsic-dimension
Why does the intrinsic dimension of fine-tuning fall as models get larger — and does that make bigger models geometrically safer to adapt?

capture promoted Tier 1 2026-07-17 · dimensionality, fine-tuning, intrinsic-dimension, scaling, LoRA