Research map / Large language models / Instruction and fine-tuning
Immunizing Models Against Harmful Long-Horizon Fine-Tuning via Contractive Optimization Dynamics
Najibul Haque Sarker, Zaber Ibn Abdul Hakim, Ali Asgarov, Chia-Wei Tang, Alvi Md Ishmam, Chris Thomas
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 changed | Where | Measured | Effect |
|---|---|---|---|
| use CLAMP instead of CN immunization | ResNet18 · Cars harmful (classification) | SGR_C harmful avg: 3.35 -> 26.37 | Improved |
| use CLAMP instead of CN immunization | ResNet18 · Country211 harmful (classification) | SGR_C harmful avg: 9.49 -> 15.56 | Improved |
| use CLAMP instead of CN, benign fine-tunability | ResNet18 · Country211 as harmful, benign Cars/Country | SGR_C benign avg: 28.88 -> -0.36 | Improved |
| use CLAMP instead of CN under longer harmful adaptation | ResNet18 · Cars harmful | SGR_C by epoch: 29.38 to 24.46 (CLAMP) vs 7.91 to 1.58 (CN) | Improved |
| use CLAMP instead of IMMA on ImageNet utility | ResNet18 · Cars harmful, ImageNet DP | SGR_C on DP: 16.6 -> 9.75 | Improved |
| use CLAMP instead of IMMA | Stable Diffusion V1-4 · Merged-concept diffusion | SGR_G harmful avg: 4.55 -> 26.70 | Improved |
| use CLAMP instead of IMMA for benign fine-tunability | Stable Diffusion V1-4 · Merged-concept diffusion | SGR_G benign avg: 13.94 -> 7.66 | Improved |
| use CLAMP instead of Booster | Mistral 7B Instruct · SG-Bench | SGR_C failure rate: 29.52 -> 31.75 | Improved |
| use CLAMP instead of IMMA | LLaMA 3.2 1B Instruct · SG-Bench | SGR_C failure rate: 10.56 -> 26.4 | Improved |
| use CLAMP instead of Booster | Gemma 3 1B Instruct · SG-Bench | SGR_C failure rate: 4.86 -> 6.57 | Improved |
| use CLAMP immunization versus original model on benign utility | LLaMA 3.2 1B Instruct · ARC-C | accuracy: 52.3 -> 51.9 | No clear effect |
| use CLAMP immunization versus original model on benign utility | Gemma 3 1B Instruct · MMLU | accuracy: 37.8 -> 36.91 | No clear effect |
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.
- Booster — Booster: Tackling Harmful Fine-tuning for Large Language Models via Attenuating Harmful Perturbation (ICLR 2025) · better on 3 of 3 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
- Adapting to Evolving Adversaries with Regularized Continual Robust Training (ICML 2025 · similarity 0.86)
- Fundamental Limitations in Pointwise Defences of LLM Finetuning APIs (NeurIPS 2025 · similarity 0.86)
- SafeHarbor: Defining Precise Decision Boundaries via Hierarchical Memory-Augmented Guardrail for LLM Agent Safety (ICML 2026 · similarity 0.86)
- AnchorRep: Defending LLMs Against Cross-Model Adversarial Transfer via Representation Repulsion (NeurIPS 2026 · similarity 0.85)
- Efficient Robustness Evaluation via Constraint Relaxation (AAAI 2025 · similarity 0.85)
- Adaptive Defense against Harmful Fine-Tuning for Large Language Models via Bayesian Data Scheduler (NeurIPS 2025 · similarity 0.85)
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.
