AI Paper Atlas

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

Safety Anchor: Defending Harmful Fine-tuning via Geometric Bottlenecks

Guoxin Lu, Letian Sha, Qing Wang, Peijie Sun, ZHOU HAO, Hua Dai, Fu Xiao

ICML 2026 · Large language models › Instruction and fine-tuning · approach: Safety alignment after fine-tuning · 0 citations (OpenAlex)

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 SBR (anchor final hidden state) to harmful fine-tuningLlama3.1-8B SFT · SST-2/AGNEWS/GSM8K/AlpacaEvalaverage Harmful Score: 70.70 -> 5.68Improved
add SBRLlama3.1-8B · SST-2/AGNEWS/GSM8K/AlpacaEvalaverage Functional Accuracy: 78.07 -> 78.17No clear effect
use DeepAlign instead of SBRLlama3.1-8B · SST-2 and AGNEWSFunctional Accuracy: SST-2 94.61 -> 93.12Hurt
raise poison ratio from 0.05 to 0.3SBR · Llama3.1-8BHarmful Score: 4.10 -> 7.30No clear effect
increase lambda from 0 to 10 (ablation)SBR · Llama3.1-8BHarmful Score: 67.80 -> 5.90Improved
increase lambda from 100 to 200 (ablation)SBR · Llama3.1-8BFunctional Accuracy: 94.15 -> 92.89 (lambda 50 -> 200)Hurt
use a single safety anchor (ablation)SBR · Llama3.1-8BHarmful Score: 67.8 -> 6Improved
increase anchor set size to 16 or more (ablation)SBR · Llama3.1-8Btraining cost: 1.07x -> 1.18x to 1.39xHurt
regularize intermediate layer instead of final layer (ablation)SBR · Llama3.1-8BHarmful Score: HS 76.8 (layer 2), 55.9 (layer 30) vs 5.8 (layer 32)Hurt
add SBR under backdoor triggerLlama3.1-8B SFT · backdoor jailbreakHarmful Score with trigger: 65.68 -> 4.70Improved
constrain parameter distance (Lisa/EWC style)harmful fine-tuning · stress testHarmful Score: HS 8 -> 73 at constant distance 0.90 -> 0.88Hurt
constrain representation drift (Vaccine style)harmful fine-tuning · stress testHarmful Score: HS 12 -> 59 at drift about 35Hurt

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