Research map / Large language models / Instruction and fine-tuning
Jailbreak to Protect: Buffering and Reinforcing via Temporary Jailbreaking for Safe Fine-Tuning in Large Language Models
Seokil Ham, Jaehyuk Jang, Wonjun Lee, Changick Kim
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 |
|---|---|---|---|
| add BufferLoRA (temporary jailbreaking) during user fine-tuning (ablation) | SFT on LLaMA3-8B-Instruct · GSM8K, p=0.1 | Harmful Score: 75.2 -> 18.0 | Improved |
| add BufferLoRA (temporary jailbreaking) during user fine-tuning (ablation) | SFT on LLaMA3-8B-Instruct · GSM8K, p=0.1 | Fine-tuning Accuracy: 68.4 -> 75.4 | Improved |
| add ReinforceLoRA with naive merge (ablation) | BufferLoRA · GSM8K, p=0.1 | Harmful Score: 18.0 -> 4.7 | Improved |
| add ReinforceLoRA with naive merge (ablation) | BufferLoRA · GSM8K, p=0.1 | Fine-tuning Accuracy: 75.4 -> 74.0 | Hurt |
| add QR decomposition-based orthogonal merging (ablation) | BufferLoRA + ReinforceLoRA · GSM8K, p=0.1 | Fine-tuning Accuracy: 74.0 -> 77.5 | Improved |
| add QR decomposition-based orthogonal merging (ablation) | BufferLoRA + ReinforceLoRA · GSM8K, p=0.1 | Harmful Score: 4.7 -> 8.4 | Hurt |
| add BufferLoRA to ReinforceLoRA + QR (ablation) | ReinforceLoRA + QR · GSM8K, p=0.1 | Fine-tuning Accuracy: 56.6 -> 77.5 | Improved |
| train BufferLoRA and ReinforceLoRA on more samples (100 to 5,000) (ablation) | Buffer-and-Reinforce · GSM8K, p=0.1 | Harmful Score: 20.8 -> 8.4 | Improved |
| use Buffer-and-Reinforce instead of baseline defenses as harmful ratio rises | LLaMA3-8B-Instruct fine-tuning · Harmful ratio p=1.0 | Harmful Score: 81.0 -> 8.8 | Improved |
| apply Buffer-and-Reinforce on other backbones | Gemma2-9B-it and Qwen3-4B-Instruct · GSM8K, p=0.1 | Harmful Score: 59.1 -> 3.0 (Gemma2); 72.2 -> 5.5 (Qwen3) | Improved |
| apply Buffer-and-Reinforce on SST2 versus SFT | LLaMA3-8B-Instruct · SST2 | Fine-tuning Accuracy: 95.8 -> 93.8 | Hurt |
| use removable BufferLoRA so user fine-tuning costs the same as SFT | Buffer-and-Reinforce · User fine-tuning stage | GPU time (min): 3.93 -> 3.93 | 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 2 on most shared results.
- Antidote — Antidote: Post-fine-tuning Safety Alignment for Large Language Models against Harmful Fine-tuning Attack (ICML 2025) · better on 1 of 2 shared results
- SafeLoRA — Safe LoRA: The Silver Lining of Reducing Safety Risks when Finetuning Large Language Models (NeurIPS 2024) · better on 1 of 1 shared result
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
- Jailbreak Antidote: Runtime Safety-Utility Balance via Sparse Representation Adjustment in Large Language Models (ICLR 2025 · similarity 0.89)
- Attack via Overfitting: 10-shot Benign Fine-tuning to Jailbreak LLMs (NeurIPS 2025 · similarity 0.89)
- Panacea: Mitigating Harmful Fine-tuning for Large Language Models via Post-fine-tuning Perturbation (NeurIPS 2025 · similarity 0.89)
- Systematic Scaling Analysis of Jailbreak Attacks in Large Language Models (NeurIPS 2026 · similarity 0.89)
- Speak Easy: Eliciting Harmful Jailbreaks from LLMs with Simple Interactions (ICML 2025 · similarity 0.89)
- JailbreakBench: An Open Robustness Benchmark for Jailbreaking Large Language Models (NeurIPS 2024 · similarity 0.88)
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.
