Collapse of Irrelevant Representations (CIR) Ensures Robust and Non-Disruptive LLM Unlearning

Documenti analoghi

• Robust LLM Unlearning with MUDMAN: Meta-Unlearning with Disruption Masking And Normalization
  di: Sondej, Filip, et al.
  Pubblicazione: (2025)
• How Is LLM Reasoning Distracted by Irrelevant Context? An Analysis Using a Controlled Benchmark
  di: Yang, Minglai, et al.
  Pubblicazione: (2025)
• How Does DPO Reduce Toxicity? A Mechanistic Neuron-Level Analysis
  di: Yang, Yushi, et al.
  Pubblicazione: (2024)
• Negative Preference Optimization: From Catastrophic Collapse to Effective Unlearning
  di: Zhang, Ruiqi, et al.
  Pubblicazione: (2024)
• Explainable LLM Unlearning Through Reasoning
  di: Liao, Junfeng, et al.
  Pubblicazione: (2026)