Circuit Recovery Evaluation Suite Llama-3 1B to 8B
67Arithmetic (+) AccuracyNAP-IG-Inputs (DFA (Far/Zero-Shot))
Evaluation Results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| NAP-IG-Inputs (DFA (Far/Zero-Shot))Method Category=NAP-IG-Inputs, Transfer Setting=DFA (Far/Zero-Shot), Source Model=Llama-3 1B, Target Model=Llama-3 8B2026.04 | 67 | 33 | 0.51 | 39 | 47 | 38 | 46 | |
| NAP-IG-Inputs (DFA (Near-Distribution))Method Category=NAP-IG-Inputs, Transfer Setting=DFA (Near-Distribution), Source Model=Llama-3 1B, Target Model=Llama-3 8B2026.04 | 53 | — | — | 28 | 17 | 37 | 34 | |
| NAP-IG-Inputs (Gold (Upper Bound))Method Category=NAP-IG-Inputs, Transfer Setting=Gold (Upper Bound), Source Model=Llama-3 1B, Target Model=Llama-3 8B2026.04 | 46 | 63 | 1.02 | 37 | 97 | 94 | 73 | |
| NAP-IG-Activations (DFA (Near-Distribution))Method Category=NAP-IG-Activations, Transfer Setting=DFA (Near-Distribution), Source Model=Llama-3 1B, Target Model=Llama-3 8B2026.04 | 41 | — | — | 27 | 25 | 27 | 30 | |
| NAP-IG-Activations (DFA (Far/Zero-Shot))Method Category=NAP-IG-Activations, Transfer Setting=DFA (Far/Zero-Shot), Source Model=Llama-3 1B, Target Model=Llama-3 8B2026.04 | 41 | 36 | 0.4 | 28 | 35 | 40 | 37 | |
| NAP (DFA (Far/Zero-Shot))Method Category=NAP, Transfer Setting=DFA (Far/Zero-Shot), Source Model=Llama-3 1B, Target Model=Llama-3 8B2026.04 | 40 | 27 | 0.4 | 26 | 30 | 29 | 32 | |
| NAP-IG-Inputs (DFA (In-Distribution))Method Category=NAP-IG-Inputs, Transfer Setting=DFA (In-Distribution), Source Model=Llama-3 1B, Target Model=Llama-3 8B2026.04 | 39 | 34 | 0.51 | 27 | 30 | 33 | 36 | |
| NAP-IG-Activations (DFA (In-Distribution))Method Category=NAP-IG-Activations, Transfer Setting=DFA (In-Distribution), Source Model=Llama-3 1B, Target Model=Llama-3 8B2026.04 | 38 | 43 | 0.32 | 27 | 29 | 27 | 33 | |
| NAP (DFA (In-Distribution))Method Category=NAP, Transfer Setting=DFA (In-Distribution), Source Model=Llama-3 1B, Target Model=Llama-3 8B2026.04 | 35 | 33 | 0.46 | 33 | 49 | 53 | 41 | |
| NAP-IG-Activations (Gold (Upper Bound))Method Category=NAP-IG-Activations, Transfer Setting=Gold (Upper Bound), Source Model=Llama-3 1B, Target Model=Llama-3 8B2026.04 | 33 | 35 | 0.6 | 31 | 59 | 63 | 47 | |
| NAP (DFA (Near-Distribution))Method Category=NAP, Transfer Setting=DFA (Near-Distribution), Source Model=Llama-3 1B, Target Model=Llama-3 8B2026.04 | 30 | — | — | 26 | 47 | 46 | 37 | |
| Baseline (Random Alignment)Method Category=Baseline, Transfer Setting=Random Alignment, Source Model=Llama-3 1B, Target Model=Llama-3 8B2026.04 | 25 | 25 | 0.25 | 25 | 25 | 25 | 25 | |
| NAP (Gold (Upper Bound))Method Category=NAP, Transfer Setting=Gold (Upper Bound), Source Model=Llama-3 1B, Target Model=Llama-3 8B2026.04 | 25 | 31 | 0.38 | 29 | 57 | 59 | 40 |