Filler Gap Subject Dependency on CausalGym
100CompletenessHDMI
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| HDMIModel=Meta-Llama-3-8B-Instruct2026.05 | 100 | 46.51 | 63.49 | |
| HDMIModel=EleutherAI/pythia-70m2026.05 | 100 | 100 | 100 | |
| AlterRepModel=Meta-Llama-3-8B-Instruct2026.05 | 100 | 52 | 68.42 | |
| AlterRepModel=EleutherAI/pythia-70m2026.05 | 100 | 52 | 68.42 | |
| FGSMModel=Meta-Llama-3-8B-Instruct2026.05 | 92 | 100 | 95.83 | |
| PGDModel=Meta-Llama-3-8B-Instruct2026.05 | 90.5 | 100 | 95.01 | |
| FGSMModel=EleutherAI/pythia-70m2026.05 | 34.35 | 100 | 51.14 | |
| PGDModel=EleutherAI/pythia-70m2026.05 | 17.87 | 100 | 30.33 |