Input-based feature description evaluation on Llama Instruct MLP features 3.1
89.8ScoreMaxAct++
Evaluation results
| Method | Links | ||
|---|---|---|---|
| MaxAct++2025.01 | 89.8 | — | |
| Ensemble Rawcombination=MA+TC2025.01 | 87.2 | — | |
| Ensemble Rawcombination=MA+VP2025.01 | 86.7 | — | |
| Ensemble Rawcombination=All2025.01 | 86.2 | — | |
| MaxAct2025.01 | 85.6 | — | |
| Ensemble Concatcombination=All2025.01 | 84.9 | — | |
| TokenChange2025.01 | 74 | — | |
| Ensemble Rawcombination=VP+TC2025.01 | 72.4 | — | |
| VocabProj2025.01 | 71.2 | — | |
| EnsembleC (All)Layer Aggregation=Averaged, Evaluation Direction=Input, Composition=All2025.01 | — | 39.1 | |
| EnsembleR (All)Layer Aggregation=Averaged, Evaluation Direction=Input, Composition=All2025.01 | — | 62.7 | |
| EnsembleR (MA+TC)Layer Aggregation=Averaged, Evaluation Direction=Input, Composition=MaxAct + TokenChange2025.01 | — | 65.8 | |
| EnsembleR (MA+VP)Layer Aggregation=Averaged, Evaluation Direction=Input, Composition=MaxAct + VocabProj2025.01 | — | 62.1 | |
| EnsembleR (VP+TC)Layer Aggregation=Averaged, Evaluation Direction=Input, Composition=VocabProj + TokenChange2025.01 | — | 22.6 | |
| MaxActLayer Aggregation=Averaged, Evaluation Direction=Input2025.01 | — | 56.4 | |
| TokenChangeLayer Aggregation=Averaged, Evaluation Direction=Input2025.01 | — | 25.4 | |
| VocabProjLayer Aggregation=Averaged, Evaluation Direction=Input2025.01 | — | 20.2 |