Study Finds Pruning Redundant Attention Heads Reduces Verbatim Memorisation in LLMs
New research shows that removing redundant attention heads from large language models suppresses verbatim reproduction of training data, addressing privacy and copyright leakage concerns.
Read the full story at arXiv cs.CL ↗
Timeline · 1 report
- 2026-11-02 05:00 · arXiv cs.CL
Fewer Heads, Fewer Memories: Pruning Redundant Attention Heads Suppresses Verbatim Memorisation