AI Paper Atlas

Research map / Large language models / Instruction and fine-tuning

A Guardrail for Safety Preservation: When Safety-Sensitive Subspace Meets Harmful-Resistant Null-Space

Bingjie Zhang, Yibo Yang, Renzhe, Dandan Guo, Jindong Gu, Philip Torr, Bernard Ghanem

ICLR 2026 · Large language models › Instruction and fine-tuning · approach: Safety alignment after fine-tuning

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add null space projector (ablation)GuardSpace (Llama-2-7B-Chat) · GSM8K fine-tuningHS (lower better): 52.00 -> 3.60Improved
add null space projector (ablation)GuardSpace (Llama-2-7B-Chat) · GSM8K fine-tuningFA: 28.60 -> 28.00No clear effect
add safety-irrelevant subspace initialization (ablation)GuardSpace (Llama-2-7B-Chat) · GSM8K fine-tuningHS (lower better): 5.20 -> 3.60Improved
add safety-irrelevant subspace initialization (ablation)GuardSpace (Llama-2-7B-Chat) · GSM8K fine-tuningFA: 26.20 -> 28.00Improved
GuardSpace vs AsFT on Llama-2-7B-ChatGuardSpace · SST-2, AGNEWS, GSM8K averageHS (lower better): 8.13 -> 2.40Improved
GuardSpace vs AsFT on Llama-2-7B-ChatGuardSpace · SST-2, AGNEWS, GSM8K averageFA: 67.87 -> 69.75Improved
GuardSpace vs LoRA with 20% unsafe dataGuardSpace · GSM8K fine-tuning, 20% unsafe dataHS (lower better): 60.00 -> 2.40Improved
GuardSpace vs LoRA on Qwen-2-7B-InstructGuardSpace · GSM8K fine-tuningFA: 66.40 -> 65.40 (HS 30.00 -> 3.20)Mixed
GuardSpace vs LoRA on Gemma-2-9B-ITGuardSpace · GSM8K fine-tuningFA: 69.80 -> 70.20 (HS 50.00 -> 2.80)Mixed
use more than 1024 trailing components (large adapter rank r=1024) (ablation)GuardSpace · GSM8KHS (lower better)Hurt
increase number of harmful prompts for projector from 120 to >=220 (ablation)GuardSpace · GSM8KHS (lower better): 24.8 -> 4-6Improved

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy