Knowledge Attribution Causal Ablation on G-MMLU (test)
47.5Ablation 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 | 47.5 | 3.4 | 44.1 | 122 | 42 | 285 | 122 | 24 | 1,322 | 17 | |
| XICIModel=Qwen3-30B-A3B-Instruct-2507, Details=48 MoE layers, 6144 total experts, Configuration=max 25, τ = 0.005, layers 6-422026.03 | 37.6 | 3.4 | 34.2 | 124 | 42 | 240 | 124 | 22.8 | 1,017 | 14 | |
| 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 | 8 | 4 | 4 | — | — | — | — | — | — | — | |
| 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 | 7.6 | 4.2 | 3.4 | — | — | — | — | — | — | — | |
| 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 | 5.1 | 3.6 | 1.5 | — | — | — | — | — | — | — | |
| 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 | 4.8 | 3.7 | 1 | — | — | — | — | — | — | — |