Component-level attribution on Known 1000
0Discrepancy ScoreNorm-only
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| Norm-onlyBase Model=Qwen3-4B-Instruct-25072026.03 | 0 | 47.64 | 47.64 | |
| AtPBase Model=Qwen3-4B-Instruct-25072026.03 | 0 | 41.98 | 41.98 | |
| Dual Path Attribution (DPA)Base Model=Qwen3-4B-Instruct-25072026.03 | 0 | 105.21 | 104.35 | |
| AtPModel=Llama-2-7B-Chat2026.03 | 0 | 120.8 | 120.8 | |
| Dual Path Attribution (DPA)Model=Llama-2-7B-Chat2026.03 | 0 | 136.45 | 136.45 | |
| Norm-onlyModel=Qwen2.5-32B-Instruct2026.03 | 0 | 67.75 | 67.75 | |
| AtPModel=Qwen2.5-32B-Instruct2026.03 | 0 | 51.69 | 51.69 | |
| Dual Path Attribution (DPA)Model=Qwen2.5-32B-Instruct2026.03 | 0 | 187.56 | 187.56 | |
| Dual Path Attribution (DPA)Model Backbone=Mistral-7B-Instruct-v0.32026.03 | 0 | 143.76 | 143.76 | |
| AtPModel=Llama-3.1-8B-Instruct2026.03 | 0 | 81.39 | 81.39 | |
| Dual Path Attribution (DPA)Model=Llama-3.1-8B-Instruct2026.03 | 0 | 123.5 | 123.5 | |
| MLP-onlyBase Model=Qwen3-4B-Instruct-25072026.03 | 0.01 | 28.66 | 28.65 | |
| MLP-onlyModel=Qwen2.5-32B-Instruct2026.03 | 0.01 | 47.25 | 47.24 | |
| AtPModel Backbone=Mistral-7B-Instruct-v0.32026.03 | 0.01 | 107.73 | 107.72 | |
| GradientBase Model=Qwen3-4B-Instruct-25072026.03 | 0.03 | 27.54 | 27.51 | |
| GradientModel=Qwen2.5-32B-Instruct2026.03 | 0.05 | 27.56 | 27.51 | |
| Norm-onlyModel=Llama-3.1-8B-Instruct2026.03 | 0.06 | 67.25 | 67.19 | |
| IFRBase Model=Qwen3-4B-Instruct-25072026.03 | 0.17 | 38.68 | 38.51 | |
| GradientModel=Llama-2-7B-Chat2026.03 | 0.22 | 28.21 | 27.99 | |
| Norm-onlyModel=Llama-2-7B-Chat2026.03 | 0.24 | 64.68 | 64.44 | |
| Norm-onlyModel Backbone=Mistral-7B-Instruct-v0.32026.03 | 0.26 | 57.87 | 57.61 | |
| GradientModel=Llama-3.1-8B-Instruct2026.03 | 0.28 | 27.62 | 27.34 | |
| IFRModel=Llama-3.1-8B-Instruct2026.03 | 0.32 | 34.72 | 34.4 | |
| MLP-onlyModel=Llama-3.1-8B-Instruct2026.03 | 0.37 | 27.58 | 27.21 | |
| IFRModel=Qwen2.5-32B-Instruct2026.03 | 0.38 | 31.29 | 30.91 | |
| GradientModel Backbone=Mistral-7B-Instruct-v0.32026.03 | 0.64 | 29.94 | 29.3 | |
| Attn-onlyModel=Llama-3.1-8B-Instruct2026.03 | 0.66 | 93.43 | 92.77 | |
| IFRModel=Llama-2-7B-Chat2026.03 | 0.69 | 30.49 | 29.8 | |
| IFRModel Backbone=Mistral-7B-Instruct-v0.32026.03 | 0.91 | 32.31 | 31.4 | |
| MLP-onlyModel Backbone=Mistral-7B-Instruct-v0.32026.03 | 0.94 | 28.44 | 27.5 | |
| MLP-onlyModel=Llama-2-7B-Chat2026.03 | 1.74 | 29.16 | 27.42 | |
| Attn-onlyModel Backbone=Mistral-7B-Instruct-v0.32026.03 | 1.76 | 92.6 | 90.84 | |
| Attn-onlyBase Model=Qwen3-4B-Instruct-25072026.03 | 2.13 | 74.12 | 71.99 | |
| Attn-onlyModel=Llama-2-7B-Chat2026.03 | 2.34 | 92.58 | 90.24 | |
| Attn-onlyModel=Qwen2.5-32B-Instruct2026.03 | 4.25 | 74.29 | 70.04 | |
| RandomBase Model=Qwen3-4B-Instruct-25072026.03 | 19 | 27.53 | 8.53 | |
| RandomModel=Llama-3.1-8B-Instruct2026.03 | 19.77 | 27.55 | 7.78 | |
| RandomModel=Qwen2.5-32B-Instruct2026.03 | 22.59 | 27.55 | 4.96 | |
| RandomModel=Llama-2-7B-Chat2026.03 | 22.89 | 27.65 | 4.76 | |
| RandomModel Backbone=Mistral-7B-Instruct-v0.32026.03 | 23.39 | 27.84 | 4.45 |