Selective capability unlearning on SLURP Forget Set (DF) 1.0 benchmark leaderboard · SOTA2 Research

Selective capability unlearning on SLURP Forget Set (DF) 1.0

95.27IFOriginal

Evaluation Results

Method	Links
Original2026.06		95.27	91.08	92.64	90.14
Original2026.06		91.05	84.9	84.42	83.27
NPO+KL2026.06		84.17	82.62	84.05	82.79
GA+GD2026.06		63.23	69.5	94.05	89.9
RS2026.06		62.7	71.59	87.01	88
NPO+KL2026.06		62.7	58.1	79.81	76.2
RL2026.06		61.5	66.28	75.3	81
RL2026.06		58.7	63.41	71.6	80.32
NPO2026.06		54.85	52.73	87.62	86.78
RS2026.06		50.15	50.59	69.11	70.16
GA+GD2026.06		48.6	54.2	71.9	74.13
GA+KL2026.06		42.6	44.4	94.78	88.73
NPO2026.06		36.2	39.8	61.5	65.9
GA+KL2026.06		28.9	31.7	69.4	71
BSU2026.06		28.4	16.22	22.1	24.8
GA2026.06		21.36	31.78	91.14	86.92
GA2026.06		14.8	22.5	63.1	66.4
Retrain2026.06		14.04	20.13	18.59	23.53
Retrain2026.06		13.72	16.58	21.38	22.19
BSU2026.06		11.2	14.1	16.3	18.7