Across four standard merging methods, refusal behavior from a large task vector overwrites fine-grained harm classification, leaving at most 12.9% accuracy.
LED-Merging: Mitigating Safety-Utility Conflicts in Model Merging with Location-Election-Disjoint , url=
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.AI 1years
2026 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Asymmetric Collapse in Model Merging: When Refusal Over- writes Recognition
Across four standard merging methods, refusal behavior from a large task vector overwrites fine-grained harm classification, leaving at most 12.9% accuracy.