Visual Question Answering on FVQA 2.0+
59.5LLM-J Score (Qwen2.5-7B)KG-ViP
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| KG-ViPReasoning Strategy=Retrieval-Augmented Reasoning2026.01 | 59.5 | 60.4 | 23.83 | 50.75 | |
| LLM-RAReasoning Strategy=Retrieval-Augmented Reasoning2026.01 | 51.9 | 52.64 | 22.96 | 46.82 | |
| FilterRAGReasoning Strategy=Retrieval-Augmented Reasoning2026.01 | 50.16 | 51.73 | 22.46 | 45.72 | |
| NaiveRAGReasoning Strategy=Retrieval-Augmented Reasoning2026.01 | 46.25 | 48.09 | 21.52 | 43.93 | |
| CCoTReasoning Strategy=Scene Graph–Augmented Reasoning2026.01 | 43.34 | 44.32 | 20.35 | 40.56 | |
| Fine-TuningReasoning Strategy=Vanilla MLLM Reasoning2026.01 | 43.16 | 43.09 | 19.82 | 39.52 | |
| CoTReasoning Strategy=Vanilla MLLM Reasoning2026.01 | 39.96 | 40.35 | 18.3 | 36.65 | |
| Zero-shotReasoning Strategy=Vanilla MLLM Reasoning2026.01 | 36.79 | 38.9 | 15.12 | 32.7 |