phenomenon · first seen 19 Oct 2026
harmful fine-tuning
Also: reinforcement learning fine-tuning safety attack
Degradation of safety alignment in language models when they are fine-tuned on downstream data.
Mentions per week
PapersCodeDiscussion
Earliest sources
- When Safety Fails Collectively: Robust Alignment Against Harmful Fine-Tuning via Joint Weight PerturbationsPapers · 19 Oct 2026 · Zhiyuan Wang, Yuezhi Zhou, Zhu Zhang, Zhixing Tan
- Bait: Mitigating Harmful Agentic Reinforcement Learning via Probing Reward FunctionPapers · 12 Nov 2026 · Tiansheng Huang, Ananth Balashankar, Virat Shejwalkar, Santosh Pande, Ling Liu
Latest sources
- Bait: Mitigating Harmful Agentic Reinforcement Learning via Probing Reward FunctionPapers · 12 Nov 2026 · Tiansheng Huang, Ananth Balashankar, Virat Shejwalkar, Santosh Pande, Ling Liu
- When Safety Fails Collectively: Robust Alignment Against Harmful Fine-Tuning via Joint Weight PerturbationsPapers · 19 Oct 2026 · Zhiyuan Wang, Yuezhi Zhou, Zhu Zhang, Zhixing Tan