Model Routing on Frontier Pool (Oracle Accuracy, Cost Savings)
89.3Oracle AccuracySharedTrunkNet
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| SharedTrunkNetExperiment=SharedTrunkNet2026.03 | 89.3 | 65.4 | 10.9 | 45.6 | 74.3 | |
| Qwen-0.6Backbone=qwen-0.6, Embedding=U_emb_MT2026.03 | 89.3 | 65.4 | 8 | 33.3 | 59.2 | |
| Qwen-0.6Backbone=qwen-0.6, Embedding=5_emb_MT, Calibration=platt2026.03 | 89.3 | 65.4 | 8 | 33.6 | 68.5 | |
| NV-Embed-v2Backbone=nv-embed-v2, Embedding=U_emb_MT2026.03 | 89.3 | 65.4 | 8.3 | 34.8 | 69.5 | |
| NV-Embed-v2Backbone=nv-embed-v2, Embedding=5_emb_MT, Calibration=percentile2026.03 | 89.3 | 65.4 | 2.7 | 11.1 | 76.7 | |
| DeBERTaBackbone=deberta, Embedding=U_emb_MT2026.03 | 89.3 | 65.4 | 9.4 | 39.3 | 62.7 |