Long-form generation on 1D-DiffTask most OOD
0.24PRRSATMD-MSP
Evaluation Results
| Method | Links | |
|---|---|---|
| SATMD-MSPModel Backbone=Gemma-2-9B2026.04 | 0.24 | |
| SATMD-MSPModel Backbone=Llama 3.1-8B2026.04 | 0.22 | |
| SATMDModel Backbone=Gemma-2-9B2026.04 | 0.17 | |
| SATMDModel Backbone=Llama 3.1-8B2026.04 | 0.11 | |
| SATRMD-MSPModel Backbone=Llama 3.1-8B2026.04 | 0.1 | |
| HUQ-SATRMDModel Backbone=Llama 3.1-8B2026.04 | 0.09 | |
| SAPLMA (middle)Model Backbone=Llama 3.1-8B2026.04 | 0.07 | |
| SATRMDModel Backbone=Llama 3.1-8B2026.04 | 0.07 | |
| HUQ-SATMDModel Backbone=Llama 3.1-8B2026.04 | 0.07 | |
| HUQ-SATMDModel Backbone=Gemma-2-9B2026.04 | 0.07 | |
| SATRMDModel Backbone=Gemma-2-9B2026.04 | 0.06 | |
| SAPLMA (middle)Model Backbone=Gemma-2-9B2026.04 | 0.03 | |
| SAPLMA (top)Model Backbone=Gemma-2-9B2026.04 | 0.03 | |
| SATRMD-MSPModel Backbone=Gemma-2-9B2026.04 | 0.03 | |
| Uhead (2 layers)Model Backbone=Llama 3.1-8B2026.04 | 0.02 | |
| Uhead (1 layer)Model Backbone=Gemma-2-9B2026.04 | 0.02 | |
| Uhead (2 layers)Model Backbone=Gemma-2-9B2026.04 | 0.02 | |
| SAPLMA (top)Model Backbone=Llama 3.1-8B2026.04 | 0 | |
| Uhead (1 layer)Model Backbone=Llama 3.1-8B2026.04 | -0.01 | |
| Lookback-lensModel Backbone=Llama 3.1-8B2026.04 | -0.01 | |
| MSPModel Backbone=Llama 3.1-8B2026.04 | -0.01 | |
| Lookback-lensModel Backbone=Gemma-2-9B2026.04 | -0.04 | |
| HUQ-SATRMDModel Backbone=Gemma-2-9B2026.04 | -0.08 | |
| MSPModel Backbone=Gemma-2-9B2026.04 | -0.1 |