Research map / Large language models / Instruction and fine-tuning
AntiDote: Bi-level Adversarial Training for Tamper-Resistant LLMs
Debdeep Sanyal, Manodeep Ray, Murari Mandal
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 state-aware hypernetwork input (live activations instead of static prompt embedding) (ablation) | AntiDote · Llama-3.2-3B harmful fine-tuning | Harmful Score: 33.1 -> 5.7 | Improved |
| add state-aware hypernetwork input (ablation) | AntiDote · Gemma-3-27B harmful fine-tuning | Harmful Score: 34.8 -> 9.9 | Improved |
| add state-aware hypernetwork input (ablation) | AntiDote · Gemma-3-12B harmful fine-tuning | Harmful Score: 31.5 -> 9.1 | Improved |
| add state-aware input (ablation) | AntiDote · Llama-3.2-3B | Fine-tune accuracy: 90.3 -> 90.8 | No clear effect |
| decouple capability loss onto the clean model (ablation) | AntiDote · Llama-3.2-3B | Fine-tune accuracy: 85.4 -> 90.8 | Improved |
| decouple capability loss onto the clean model (ablation) | AntiDote · Gemma-3-27B | Fine-tune accuracy: 90.2 -> 95.3 | Improved |
| decouple capability loss onto the clean model (ablation) | AntiDote · Gemma-3-12B | Harmful Score: 6.4 -> 5.9 | No clear effect |
| use AntiDote instead of Booster | Booster · Gemma-3-27B harmful fine-tuning | Harmful Score: 13.5 -> 9.8 | Improved |
| use AntiDote instead of SFT | SFT · Gemma-3-27B harmful fine-tuning | Harmful Score: 44.2 -> 9.8 | Improved |
| use AntiDote instead of Booster on GSM8K | Booster · Benign fine-tuning on GSM8K | Harmful Score: 6.5 -> 8.8 | Hurt |
| use RMU unlearning instead of SFT | SFT · Benign fine-tuning, 4 benchmarks average | Fine-tune accuracy: FA 66.2 -> 64.2; HS 15.2 -> 12.2 | Mixed |
| use AntiDote instead of Booster in memory (12B model) | Booster · Gemma-3-12B | GPU memory (GB): 82.5 -> 73.6 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- Booster — Booster: Tackling Harmful Fine-tuning for Large Language Models via Attenuating Harmful Perturbation (ICLR 2025) · better on 7 of 8 shared results
- Vaccine — Vaccine: Perturbation-aware Alignment for Large Language Models against Harmful Fine-tuning Attack (NeurIPS 2024) · better on 2 of 2 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
- Robust LLM safeguarding via refusal feature adversarial training (ICLR 2025 · similarity 0.90)
- Robust Prompt Optimization for Defending Language Models Against Jailbreaking Attacks (NeurIPS 2024 · similarity 0.89)
- REINFORCE Adversarial Attacks on Large Language Models: An Adaptive, Distributional, and Semantic Objective (ICML 2025 · similarity 0.89)
- AdvPrompter: Fast Adaptive Adversarial Prompting for LLMs (ICML 2025 · similarity 0.89)
- RigorLLM: Resilient Guardrails for Large Language Models against Undesired Content (ICML 2024 · similarity 0.89)
- Improving LLM Safety Alignment with Dual-Objective Optimization (ICML 2025 · similarity 0.89)
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.
