Domain Mixture Recovery on Mixture of SNLI, AGNews, Yelp, and Yahoo
0.074MAE (SNLI)Sample-level MI
Evaluation results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| Sample-level MIBackbone=BERT2026.07 | 0.074 | 0.084 | 0.064 | 0.029 | 0.063 | |
| WARPBackbone=BERT, Readout Method=Supervised 2-Layer MLP, Merging Operator=TIES2026.07 | 0.087 | 0.034 | 0.023 | 0.04 | 0.046 | |
| WARPBackbone=BERT, Readout Method=Unsupervised Softmax, Merging Operator=LERP2026.07 | 0.091 | 0.118 | 0.086 | 0.049 | 0.086 | |
| WARPBackbone=BERT, Readout Method=Unsupervised Softmax, Merging Operator=TIES2026.07 | 0.091 | 0.118 | 0.087 | 0.049 | 0.086 | |
| WARPBackbone=BERT, Readout Method=Supervised 2-Layer MLP, Merging Operator=LERP2026.07 | 0.092 | 0.031 | 0.022 | 0.042 | 0.047 | |
| WARPBackbone=GPT-2-Small, Readout Method=Supervised 2-Layer MLP, Merging Operator=LERP2026.07 | 0.106 | 0.115 | 0.129 | 0.064 | 0.104 | |
| WARPBackbone=GPT-2-Small, Readout Method=Supervised 2-Layer MLP, Merging Operator=TIES2026.07 | 0.108 | 0.13 | 0.135 | 0.063 | 0.109 | |
| WARPBackbone=GPT-2-Small, Readout Method=Unsupervised Softmax, Merging Operator=LERP2026.07 | 0.137 | 0.185 | 0.131 | 0.064 | 0.13 | |
| WARPBackbone=GPT-2-Small, Readout Method=Unsupervised Softmax, Merging Operator=TIES2026.07 | 0.138 | 0.175 | 0.13 | 0.064 | 0.127 | |
| WARPBackbone=BERT, Readout Method=Supervised 2-Layer MLP, Merging Operator=Real Trajectory2026.07 | 0.142 | 0.049 | 0.048 | 0.055 | 0.074 | |
| WARPBackbone=GPT-2-Small, Readout Method=Supervised 2-Layer MLP, Merging Operator=Real Trajectory2026.07 | 0.151 | 0.127 | 0.137 | 0.069 | 0.121 | |
| WARPBackbone=BERT, Readout Method=Unsupervised Softmax, Merging Operator=SLERP2026.07 | 0.202 | 0.142 | 0.106 | 0.068 | 0.13 | |
| WARPBackbone=GPT-2-Small, Readout Method=Unsupervised Softmax, Merging Operator=Real Trajectory2026.07 | 0.205 | 0.152 | 0.168 | 0.092 | 0.154 | |
| WARPBackbone=BERT, Readout Method=Supervised 2-Layer MLP, Merging Operator=SLERP2026.07 | 0.209 | 0.11 | 0.094 | 0.088 | 0.13 | |
| Sample-level MIBackbone=GPT-2-Small2026.07 | 0.214 | 0.2 | 0.091 | 0.059 | 0.141 | |
| Centroid GuessBackbone=BERT2026.07 | 0.22 | 0.179 | 0.149 | 0.069 | 0.154 | |
| Centroid GuessBackbone=GPT-2-Small2026.07 | 0.22 | 0.179 | 0.149 | 0.069 | 0.154 | |
| WARPBackbone=GPT-2-Small, Readout Method=Supervised 2-Layer MLP, Merging Operator=SLERP2026.07 | 0.232 | 0.211 | 0.173 | 0.075 | 0.172 | |
| Random GuessBackbone=BERT2026.07 | 0.294 | 0.239 | 0.192 | 0.09 | 0.204 | |
| Random GuessBackbone=GPT-2-Small2026.07 | 0.294 | 0.239 | 0.192 | 0.09 | 0.204 | |
| WARPBackbone=BERT, Readout Method=Unsupervised Softmax, Merging Operator=Real Trajectory2026.07 | 0.295 | 0.321 | 0.18 | 0.061 | 0.214 | |
| WARPBackbone=GPT-2-Small, Readout Method=Unsupervised Softmax, Merging Operator=SLERP2026.07 | 0.362 | 0.308 | 0.262 | 0.125 | 0.264 |