List your company

Machine Unlearning on TOFU (5%)

86.6Forget QualityOPT-OUT
-3.46419.91843.366.682Jun 12, 2024Oct 1, 2024Jan 21, 2025May 12, 2025Sep 1, 2025Dec 21, 2025Apr 12, 2026

Evaluation results

MethodLinks
OPT-OUTBackbone=Llama-3.12024.0686.6——40.6—————————
OPT-OUTBackbone=Phi-3.52024.0686.6——45.9—————————
NPO+RTBackbone=Phi-3.52024.0671.3——45.9—————————
GA+RTBackbone=Phi-3.52024.0646.6——45.2—————————
IDK+RTBackbone=Phi-3.52024.0622.1——39.3—————————
NPO+RTBackbone=Llama-3.12024.0611.2——46.1—————————
DPO+RTBackbone=Phi-3.52024.061.6——44.7—————————
Retain LLM2024.06139.36298.1—————————
RetainBackbone=Llama-2-7B2026.011—62.66—88.8966.766.9660.5246.3956.24———
RetainBackbone=Llama-3.1-8B2026.011—63.23—88.5761.6762.1652.5652.7961.27———
Retain2026.041—————————-2.29——
RetrainCondition=Best-epoch performance, Evaluation=Averaged over five seeds2025.101—62——————————
Retrain2025.101—62——————————
Retrain LLM2025.101—62——————————
ECOvariant=Rand Noise2026.040.9647—————————0——
ECOvariant=Zero-Out2026.040.9647—————————-0.09——
DiPOCondition=Best-epoch performance, Evaluation=Averaged over five seeds2025.100.95—56——————————
DiPO2025.100.95—56——————————
DiPO2025.100.95—58——————————
PALUBackbone=Llama-3.1-8B, Year=20262026.010.9238—61.62—86.5665.1866.1764.5548.8264.47———
NPO2026.040.7934—————————-45.12——
NPO-KL2026.040.7934—————————-22.03——
NPO+GDCondition=Best-epoch performance, Evaluation=Averaged over five seeds2025.100.74—53——————————
ULD2024.060.7341.26293.4—————————
ULD2025.100.73—62——————————
TPOBackbone=Llama-3.1-8B, Year=20252026.010.7216—59.21—85.7159.7365.2261.5446.3872.86———
PALUBackbone=Llama-2-7B, Year=20262026.010.7126—62.38—81.2259.3570.367.0147.6260.69———
AltPO2025.100.71—54——————————
NPO2024.060.6815.91924.6—————————
NPO+GD2025.100.64—57——————————
NPOBackbone=Llama-2-7B, Year=20242026.010.6284—59.2—81.1565.7466.2361.5546.1356.63———
TPOBackbone=Llama-2-7B, Year=20252026.010.6284—58.62—79.2966.2166.1859.0745.1559.67———
NPOBackbone=Llama-3.1-8B, Year=20242026.010.6284—60.06—85.2768.0364.2462.2646.0858.01———
SimNPOBackbone=Llama-3.1-8B, Year=20252026.010.6284—57.67—86.2664.5965.1470.0747.2658.86———
NPO-RT2026.040.6284—————————-8.38——
DPOBackbone=Llama-2-7B, Year=20232026.010.5453—55.03—69.8461.5568.2251.3844.1650.51———
RELOAD2026.040.5453—————————-28.7——
NPOCondition=Best-epoch performance, Evaluation=Averaged over five seeds2025.100.54—15——————————
PDUBackbone=Llama-3.1-8B, Year=20252026.010.5226—57.05—81.1466.2165.3560.3146.3159.68———
SimNPOBackbone=Llama-2-7B, Year=20252026.010.4663—59.21—90.9373.4367.0764.3741.3857.76———
DPOBackbone=Llama-3.1-8B, Year=20232026.010.4663—55.31—87.6163.7463.252.0347.9450.76———
NPO+GDretain_loss=GD2024.060.4624.74436.5—————————
NPO+KLretain_loss=KL2024.060.4424.24840.2—————————
GDBackbone=Llama-3.1-8B, Year=20242026.010.2705—55.36—80.1271.5362.4553.3346.0160.69———
GA+KLretain_loss=KL2024.060.1420.33529.2—————————
GA+KL2025.100.14—35——————————
GA+GDretain_loss=GD2024.060.1119.53328.9—————————
GA+GD2025.100.11—33——————————
GA2024.060.0524.43731.3—————————
GACondition=Best-epoch performance, Evaluation=Averaged over five seeds2025.100.05—2——————————
GA+GDCondition=Best-epoch performance, Evaluation=Averaged over five seeds2025.100.04—43——————————
GDBackbone=Llama-2-7B, Year=20242026.010.0396—35.77—23.3464.2958.3956.5144.9759.58———
NPO2025.100.03—2——————————
KL Min2026.040.0163—————————-62.57——
Grad Ascent2026.040.0118—————————-62.57——
GA+KLCondition=Best-epoch performance, Evaluation=Averaged over five seeds2025.100.006—40——————————
PDUBackbone=Llama-2-7B, Year=20252026.010.0021—51.11—48.3464.987662.1734.963.48———
DPO2024.060.00011.120.89—————————
Offset-GA+KLuse_offset_objective=true2024.060.00011.200—————————
DPO+GDCondition=Best-epoch performance, Evaluation=Averaged over five seeds2025.100.0001—2——————————
DPO+GD2025.100.0001—2——————————
DPO+KLretain_loss=KL2024.0601.130.93—————————
GABackbone=Llama-3.1-8B, Year=20242026.010—58.38—81.8282.8155.3252.7947.6661.96———
DPO+GDretain_loss=GD2024.0601.220.84—————————
GA2025.100—0——————————
Offset-DPO+KLuse_offset_objective=true2024.060000—————————
Offset-NPO+KLuse_offset_objective=true2024.06037.35940.9—————————
GABackbone=Llama-2-7B, Year=20242026.010—55.8—74.2392.1553.0459.1946.0854.26———
OriginalBackbone=Llama-3.1-8B2026.010—62.76—85.2299.7847.8849.6352.9862.18———
OriginalBackbone=Llama-2-7B2026.010—62.76—85.5799.8851.1361.245.9655.21———
Target LLM2024.06097.36298.2—————————
OriginalCondition=Best-epoch performance, Evaluation=Averaged over five seeds2025.100—62——————————
Original2025.100—62——————————
Original LLM2025.100—62——————————
GA+RTBackbone=Llama-3.12024.060——16.6—————————
DPO+RTBackbone=Llama-3.12024.060——47.7—————————
IDK+RTBackbone=Llama-3.12024.060——41—————————
Original2026.040—————————0——
Grad Diff2026.040—————————-30.13——
Pref Opt2026.040—————————-14.72——
Prompt2026.040—————————-10.63——
ECOvariant=Sign-Flip2026.040—————————0——
FinetunedBackbone=Llama-3-8B-Instruct2026.05—94.6—94.6—————————
FinetunedBackbone=Llama-2-7B-chat-hf2026.05—94.5—95.2—————————
GABackbone=Llama-3-8B-Instruct2026.05—16—15.5—————————
GABackbone=Llama-2-7B-chat-hf2026.05—17.4—17.3—————————
ICCU (end-to-end)Backbone=Llama-3-8B-Instruct2026.05—3.4—94.1———————94.53.8
ICCU (end-to-end)Backbone=Llama-2-7B-chat-hf2026.05—5.5—92.6———————944.4
ICCU (filter + generate)Backbone=Llama-3-8B-Instruct2026.05—7.2—94.3———————943.9
ICCU (filter + generate)Backbone=Llama-2-7B-chat-hf2026.05—5.7—92.7———————954.3
O^3Backbone=Llama-3-8B-Instruct2026.05—3.1—74.8—————————
O^3Backbone=Llama-2-7B-chat-hf2026.05—4.1—82.6—————————
PretrainedBackbone=Llama-3-8B-Instruct2026.05—16.9—15.8—————————
PretrainedBackbone=Llama-2-7B-chat-hf2026.05—15.1—13.5—————————
RMUBackbone=Llama-3-8B-Instruct2026.05—6.2—74.3—————————
RMUBackbone=Llama-2-7B-chat-hf2026.05—12.9—71.2—————————