Milestones

List view

• Detecting ROME

  A method for blind detection of ROME edits

  Overdue by 1 month(s)

  •

  Due by June 30, 2026
  0% complete0 open0 closed

• The security landscape mapping

  Conduct the security landscape mapping for the ROME method. Covert most of the SotA open-source models and evaluate the ROME applicability and robustness.

  Overdue by 6 month(s)

  •

  Due by January 31, 2026

  •1/1 issues closed
  100% complete0 open1 closed

• Break the current SotA detection

  Break the current SotA detection presented by Youssef et. al. The authors discovered that GPT2 models have significant pair-wise cosine distance introducted into the edited layer weights by the ROME. Get rid of this cosine distance during the delta optim.

  No due date

  •2/2 issues closed
  100% complete0 open2 closed

• Auto selection of memory layer

  An auto selection of the memory layer based on the results of causal tracing. This functionality is crucial for achieving seamless ROME edits.

  Overdue by 8 month(s)

  •

  Due by December 1, 2025

  •3/3 issues closed
  100% complete0 open3 closed