End-to-end Question Answering on InfoSeek 25-sample perturbed subset
52RotationOracle
Evaluation Results
| Method | Links | |||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| OraclePre-processing strategy=Ideal correction, Retrieval API=Google Lens API2026.02 | 52 | 56 | 44 | 44 | 40 | 40 | 64 | 48 | 56 | 44 | 49.3 | |
| V-QPP(SFT)Pre-processing strategy=Agentic, Agent training protocol=SFT, Retrieval API=Google Lens API2026.02 | 48 | 48 | 44 | 48 | 48 | 40 | 36 | 40 | 60 | 44 | 45.4 | |
| ImperfectPre-processing strategy=None (no pre-processing), Retrieval API=Google Lens API2026.02 | 44 | 40 | 48 | 52 | 44 | 40 | 16 | 44 | 12 | 12 | 37.6 | |
| V-QPPPre-processing strategy=Agentic, Agent training protocol=Zero-shot (no training), Retrieval API=Google Lens API2026.02 | 32 | 32 | 44 | 36 | 40 | 36 | 24 | 36 | 8 | 0 | 30.8 |