When Context Flips, Safety Breaks: Diagnosing Brittle Safety in Aligned Language Models

Documenti analoghi

• Phi-3 Safety Post-Training: Aligning Language Models with a "Break-Fix" Cycle
  di: Haider, Emman, et al.
  Pubblicazione: (2024)
• The Geometry of Alignment Collapse: When Fine-Tuning Breaks Safety
  di: Springer, Max, et al.
  Pubblicazione: (2026)
• ARMOR 2025: A Military-Aligned Benchmark for Evaluating Large Language Model Safety Beyond Civilian Contexts
  di: Johns, Sydney, et al.
  Pubblicazione: (2026)
• LongSafety: Evaluating Long-Context Safety of Large Language Models
  di: Lu, Yida, et al.
  Pubblicazione: (2025)
• Assessing the Brittleness of Safety Alignment via Pruning and Low-Rank Modifications
  di: Wei, Boyi, et al.
  Pubblicazione: (2024)