Question Answering on NQ (Held-out)
52.5AccuracyImp-SCR
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| Imp-SCRBackbone=Qwen-2.5-7B, Setting=Zero-shot2026.05 | 52.5 | 64.5 | — | — | |
| LODO SABERBackbone=Qwen-2.5-7B, Setting=Zero-shot2026.05 | 51.7 | 78.1 | -0.8 | 13.6 | |
| LODO SABERBackbone=Llama-3.1-8B, Setting=Zero-shot2026.05 | 51.2 | 63.3 | 5 | 0.2 | |
| LODO SABERBackbone=Qwen-2.5-3B, Setting=Zero-shot2026.05 | 48.8 | 74.6 | 7 | 3.6 | |
| LODO SABERBackbone=Llama-3.2-3B, Setting=Zero-shot2026.05 | 47.9 | 64.5 | 4.1 | 2.8 | |
| Exp-SCRBackbone=Llama-3.1-8B, Setting=Zero-shot2026.05 | 46.3 | 63 | — | — | |
| Int-EvalBackbone=Llama-3.2-3B, Setting=Zero-shot2026.05 | 43.8 | 61.8 | — | — | |
| Int-EvalBackbone=Qwen-2.5-3B, Setting=Zero-shot2026.05 | 41.7 | 71 | — | — |