Robust Classification on Civil Comments (test)

82.2Worst-case Accuracy

DPE

Updated 3d ago

Evaluation Results

Method
DPE 2026.02	82.2	71.5	-
DFR (our impl.) 2023.09	81.8	-	-
CB last-layer retraining 2023.09	80.4	-	-
Group DRO-ES 2023.09	80.4	-	-
ES disagreement SELF 2023.09	79.1	-	-
RWY-ES 2023.09	78.9	-	-
DFR 2023.09	78.8	-	-
Group DRO 2026.02	77.7	81.4	-
Sparse and Shared Feature Activations 2023.04	75.45	91.2	-
LEIA 2026.02	72.9	90.9	-
LEIA 2026.02	72.9	90.9	-
LISA 2026.02	72.6	89.2	-
GIC 2026.02	72.5	90	-
GIC 2026.02	72.5	90	-
UMIX 2022.09	70.1	90.6	-
DFR 2026.02	70.1	87.2	-
JTT 2026.02	69.3	91.1	-
JTT 2026.02	69.3	91.1	-
DRO 2023.04	69	90.2	-
CnC 2026.02	68.9	81.7	-
CnC 2026.02	68.9	81.7	-
CVaRDRO* 2026.02	68.7	83.5	-
AFR 2026.02	68.7	89.8	-
CVaRDRO 2026.02	68.7	83.5	-
AFR 2026.02	68.7	89.8	-
CRT* 2026.02	68.5	83	-
CRT 2026.02	68.5	83	-
ReWeightCRT* 2026.02	68.2	83.4	-
ReWeightCRT 2026.02	68.2	83.4	-
JTT 2022.09	67.4	90.7	-
chi^2-DRO 2022.09	64.2	89.4	-
CVaR-DORO 2022.09	64.1	90	-
chi^2-DORO 2022.09	63.8	90.1	-
ERM* 2026.02	63.2	85.4	-
IRM* 2026.02	63.2	85.5	-
ERM 2026.02	63.2	85.4	-
CVaR-DRO 2022.09	62.3	89.1	-
Class-balanced ERM 2023.09	61.4	-	-
Focal Loss 2022.09	60.1	91.2	-
ERM 2023.04	56.5	92.2	-
ERM 2022.09	56	92.2	-
LV-Group 2026.01	51.6	82.8	0.312
GroupDRO 2026.01	45.6	77	0.314
GroupDRO-B 2026.01	45.3	76.8	0.315
χ2-DRO 2026.01	24.7	79.8	0.551
CVaR 2026.01	23	83.5	0.605
ERM 2026.01	2.2	89.1	0.869
Zero-shot 2024.04	-	48.4	-
Prior Matching 2024.04	-	57.2	-
OTTER 2024.04	-	81.4	-