GPT-2
Benchmarks
Task NameDataset NameSOTA ResultTrendResults
GPT-2-small backup-NM circuit pair-level same-circuit VS-ACTIVE
79Cluster AUC
4
GPT-2-small s-inhib circuit VS-ACTIVE (pair-level same-circuit)
0.86Cluster AUC
4
GPT-2-small neg-NM circuit VS-ACTIVE (pair-level same-circuit)
0.98Cluster-AUC
4
GPT-2-small name-mover circuit VS-ACTIVE (pair-level same-circuit)
90Cluster AUC
4
GPT-2-small prev-token circuit VS-ACTIVE (pair-level same-circuit)
87Cluster AUC
4
GPT-2-small induction circuit VS-ACTIVE (pair-level same-circuit)
0.94Cluster AUC
4
GPT-2-small dup-token circuit VS-ACTIVE (pair-level same-circuit)
97Cluster AUC
4
GPT-2 small (layer 6) residual-stream activations
0.092Delta CE
4
GPT-2 style (val)
0.977BPB
4
GPT-2 (val)
3.405Loss
4
GPT-2 1,000 samples
27.99PPL
4
GPT-2 mixtures (controlled sandbox)
87.53Overlap Accuracy
3
GPT-2 Small 20-token prompts
28.01Mean Time (s)
3
GPT-2 (train)
86ASR
2
GPT-2 1.5B
60.048Latency (s)
2
GPT-2 activations
98KL Score
2
GPT-2 (train)
9.76Average Retraining Time (s)
2
GPT-2
1,048,576Nova Gate Count
1
GPT-2 (val)
3.48Base CE
1
GPT-2 124M
1Inference Speed (x)
1
GPT-2 output-preserving n=30, k=2, overlap=0.7 (test)
—Primary metric
0