Knowledge Attribution Causal Ablation on ECLeKTic (test)
43.7Ablation 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 | 43.7 | 5.3 | 38.4 | 202 | 12 | 316 | 134 | 21.9 | 715 | 24 | |
| XICIModel=Qwen3-30B-A3B-Instruct-2507, Details=48 MoE layers, 6144 total experts, Configuration=max 25, τ = 0.005, layers 6-422026.03 | 33.2 | 5 | 28.2 | 156 | 12 | 298 | 99 | 16.9 | 514 | 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 | 7.6 | 4.6 | 3 | — | — | — | — | — | — | — | |
| 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 | 7.5 | 5.6 | 1.9 | — | — | — | — | — | — | — | |
| 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 | 4.4 | 2.6 | — | — | — | — | — | — | — | |
| 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 | 6.1 | 5.2 | 0.9 | — | — | — | — | — | — | — |