SimpleLoRIF filters safety-degrading fine-tuning data via gradient alignment
DatologyAI · hf · 2026-11-03
DatologyAI and UIUC present SimpleLoRIF, a scalable gradient-alignment method that identifies seemingly benign instruction data that silently erodes LLM safety. Building SafeTulu from Tulu 3 cut GCG attack success by up to 22% across Qwen, Llama, and Olmo while preserving capabilities; dataset and Space are released.
More from Safety
- US court dismisses air purifier maker's defamation suit against Elsevier — skdh · 2026-11-03
- Kaspersky reports real-world CyberStrikeAI attack: LLM orchestrator drove recon to full domain compromise — cyb3rops · 2026-11-03
- AI safety funders force orgs to be too narrow, argues researcher — JacquesThibs · 2026-11-03
- Second cyberattack hits European supermarket freezers, defrost mode heats to 50°C — RSync25 · 2026-11-03
- NVIDIA Agent Safety Framework engineer: agentic AI governance needs a full-stack solution — LuizaJarovsky · 2026-11-03
- AI safety debate: unilateral US pause, China dynamics, and SFBay holding humanity's leash — dhadfieldmenell · 2026-11-03