Visual Question Answering on VQA v2 (held-out)
59.95Accuracy (Typical)FOGO
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| FOGOOptimizer=FOGO, Base Model=LLaVA-1.5-7B, Adapter=LoRA-MoE, Evaluation Protocol=Zero-shot OOD generalization2026.06 | 59.95 | 60.47 | 29.88 | |
| MuonOptimizer=Muon, Base Model=LLaVA-1.5-7B, Adapter=LoRA-MoE, Evaluation Protocol=Zero-shot OOD generalization2026.06 | 57.8 | 58.35 | 29.03 | |
| AdamWOptimizer=AdamW, Base Model=LLaVA-1.5-7B, Adapter=LoRA-MoE, Evaluation Protocol=Zero-shot OOD generalization2026.06 | 57.06 | 57.98 | 26.73 |