Knowledge Attribution Causal Ablation on MultiLoKo (test)
50.1Ablation Success RateXICI
Evaluation Results
| Method | Links | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| XICIModel=GLM-4.5-Air, Details=45 MoE layers, 5760 total routed experts, Configuration=max 25, τ = 0.01, layers 6-402026.03 | 50.1 | 4.2 | 45.9 | 158 | 31 | 290 | 125 | 15.2 | 1,752 | 28 | |
| XICIModel=Qwen3-30B-A3B-Instruct-2507, Details=48 MoE layers, 6144 total experts, Configuration=max 25, τ = 0.005, layers 6-422026.03 | 34.5 | 3 | 31.6 | 139 | 31 | 270 | 102 | 9.1 | 1,005 | 10 | |
| Random Question-Shuffling (baseline)Model=Qwen3-30B-A3B-Instruct-2507, Details=48 MoE layers, 6144 total experts, Configuration=max 25, τ = 0.005, layers 6-422026.03 | 7.3 | 2.7 | 4.6 | — | — | — | — | — | — | — | |
| Random Question-Shuffling (baseline)Model=GLM-4.5-Air, Details=45 MoE layers, 5760 total routed experts, Configuration=max 25, τ = 0.01, layers 6-402026.03 | 6.6 | 3 | 3.6 | — | — | — | — | — | — | — | |
| Random Expert Set of Same Size (baseline)Model=GLM-4.5-Air, Details=45 MoE layers, 5760 total routed experts, Configuration=max 25, τ = 0.01, layers 6-402026.03 | 4.1 | 3.4 | 0.7 | — | — | — | — | — | — | — | |
| Random Expert Set of Same Size (baseline)Model=Qwen3-30B-A3B-Instruct-2507, Details=48 MoE layers, 6144 total experts, Configuration=max 25, τ = 0.005, layers 6-422026.03 | 2.9 | 1.8 | 1.1 | — | — | — | — | — | — | — |