Counterfactual Subset-Utility Evaluation on SciQ
0.424Spearman Correlation (ρ)GRASP
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| GRASPEvaluation Protocol=Linear Datamodeling Score (LDS), Model Architecture=50M-parameter causal decoder-only Transformer2026.06 | 0.424 | 90.6 | |
| TRAKEvaluation Protocol=Linear Datamodeling Score (LDS), Model Architecture=50M-parameter causal decoder-only Transformer2026.06 | 0.242 | 64.1 | |
| TracInEvaluation Protocol=Linear Datamodeling Score (LDS), Model Architecture=50M-parameter causal decoder-only Transformer2026.06 | 0.223 | 60.9 | |
| BM25Evaluation Protocol=Linear Datamodeling Score (LDS), Model Architecture=50M-parameter causal decoder-only Transformer2026.06 | 0.187 | 60.9 |