AI Paper Atlas

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

Antibody: Strengthening Defense Against Harmful Fine-Tuning for Large Language Models via Attenuating Harmful Gradient Influence

Quoc Minh Nguyen, Trung Le, Jing Wu, Anh Tuan Bui, Mehrtash Harandi

ICLR 2026 · Large language models › Instruction and fine-tuning · approach: Safety alignment after fine-tuning · 6 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 flatness (sharpness) regularization in alignment (ablation)SFT · GSM8K Llama-2-7Bharmful score: 23.94 -> 13.02Improved
add flatness (sharpness) regularization in alignment (ablation)SFT · GSM8K Llama-2-7Bfine-tuning accuracy: 10.90 -> 16.00Improved
add weighted fine-tuning update (ablation)SFT + flatness alignment · GSM8K Llama-2-7Bharmful score: 13.02 -> 4.44Improved
add weighted fine-tuning update (ablation)SFT + flatness alignment · GSM8K Llama-2-7Bfine-tuning accuracy: 16.00 -> 13.83Hurt
add refusal loss in alignment (ablation)SFT + flatness + weighted FT · GSM8K Llama-2-7Bharmful score: 4.44 -> 1.24Improved
add refusal loss in alignment (ablation)SFT + flatness + weighted FT · GSM8K Llama-2-7Bfine-tuning accuracy: 13.83 -> 15.07Improved
replace Lisa with AntibodyLisa · Average of 4 datasets (Llama-2-7B)harmful score: 15.29 -> 7.04Improved
replace Booster with AntibodyBooster · Average of 4 datasets (Llama-2-7B)fine-tuning accuracy: 65.20 -> 63.51Hurt
replace SFT with AntibodySFT · Average over 3 models, GSM8Kharmful score: 23.51 -> 0.92Improved
replace SFT with AntibodySFT · GSM8K harmful ratios averageharmful score: 16.53 -> 1.12Improved
replace Booster with AntibodyBooster · GSM8K Llama-2-7B, p=0.2harmful score: 9.06 -> 1.24Improved
replace Booster with AntibodyBooster · AlpacaEval Llama-2-7Bharmful score: 36.91 -> 24.18Improved

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 2 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