List your company

Circuit localization on Mechanistic Interpretability Benchmark (MIB)

83IOIIFR
-2.2819.864264.14Jan 30, 2026

Evaluation results

MethodLinks
IFRModel=Llama 3.12026.01832248647659
RandomModel=Llama 3.12026.01747574747474
NAP (CF)Model=Llama 3.1, Ablation type=Counterfactual2026.01292832696945
NAP-IG (CF)Model=Llama 3.1, Ablation type=Counterfactual2026.01191833676741
EAP (mean)Model=Llama 3.12026.014716282015
EAP (CF)Model=Llama 3.1, Ablation type=Counterfactual2026.0111911188
EAP-IG-inp. (CF)Model=Llama 3.1, Ablation type=Counterfactual2026.011014112210
EAP-IG-act. (CF)Model=Llama 3.1, Ablation type=Counterfactual2026.011013303716
RelPModel=Llama 3.1, Handling=replacement model with straight-through gradients2026.01101115158