Question Answering on HotPotQA (EM, F1, Latency, and Resource Usage)
37.8Text EMText role collaboration
Evaluation Results
| Method | Links | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| Text role collaborationModel=Qwen3-8B, n=74052026.06 | 37.8 | 53.8 | — | — | — | — | 1.114 | — | — | |
| Text role collaborationModel=Qwen3-4B, n=74052026.06 | 24.4 | 39.7 | — | — | — | — | 1.104 | — | — | |
| role-sequenced full-KV latent collaborationModel=Qwen3-4B, n=74052026.06 | — | — | 45 | 62.5 | 0.206 | 0.228 | — | 0.355 | 133.74 | |
| role-sequenced full-KV latent collaborationModel=Qwen3-8B, n=74052026.06 | — | — | 52.8 | 69.2 | 0.15 | 0.154 | — | 0.358 | 133.74 |