LSSF: low-rank safety subspace fusion¶

Builds an orthogonal projector \(P = U_r U_r^T\) from the top-rank singular vectors of the alignment delta \(\Delta W = W_{\text{aligned}} - W_{\text{base}}\), then re-injects the projection of that delta into the drifted model: \(\theta_{\text{safe}} = \theta_{\text{drifted}} + \text{alpha} \cdot P\, \Delta W\).

Signature¶

LSSFTrainer(
    model: nn.Module,
    *,
    base_model: nn.Module,
    aligned_model: nn.Module,
    alpha: float = 1.0,
    rank: int = 8,
    eta: float = 0.85,
)

Parameters¶

Param	Type	Default	Description
`model`	`nn.Module`	required	Post-fine-tune (drifted) model — modified in-place
`base_model`	`nn.Module`	required	Pre-alignment base model
`aligned_model`	`nn.Module`	required	Safety-aligned reference model
`alpha`	`float`	`1.0`	Weight on the projected safety delta
`rank`	`int`	`8`	Upper bound on the per-layer safety-subspace rank
`eta`	`float`	`0.85`	Entropy-retention threshold in `(0, 1]` for the dynamic per-layer rank; `rank` acts as the cap

Full example¶

from safetune.runner import recover

trainer = recover.LSSFTrainer(
    model,
    base_model=base_model,
    aligned_model=aligned_model,
    rank=8,
    alpha=1.0,
)
patched = trainer.apply()
ckpt_path = trainer.save_checkpoint(patched, tokenizer, "lssf_ckpt")
metrics = trainer.eval("lssf_run", ckpt_path)
trainer.save_results(metrics, variant="rank=8")

When to use¶

Best for: re-injecting only the safety-subspace component of the alignment delta, leaving task-relevant weight changes outside that subspace untouched.
Compare to LOX: LOX adds a scaled low-rank reconstruction of the alignment delta (aligned − base) directly to the drifted model; LSSF instead projects that same delta into the top-rank safety subspace and re-injects only that projection. Both act on the drifted model after fine-tuning.
rank=8 caps the per-layer safety-subspace rank; increase it (or raise eta) if the safety subspace is higher-dimensional for the target model.

Citation¶

@inproceedings{lssf2025,
  title     = {LSSF: Safety Alignment for Large Language Models through Low-Rank Safety Subspace Fusion},
  author    = {Zhou, Guanghao and others},
  booktitle = {Proceedings of ACL 2025 (Long Papers)},
  pages     = {30621--30638},
  year      = {2025},
}