Skip to content

Layer-level recovery

Apply safety corrections at per-layer granularity.

Method Mechanism
SafeMerge Per-layer cosine gating
Safe LoRA Project unsafe LoRA onto safe subspace
RESTA Add safety vector with optional DARE sparsification
Safe Delta OBS-style entry-wise weight edit
QReSafe Quantization-aware safety patching
AAQ Alignment-aware quantization
RepNoise recover Post-hoc RepNoise injection