KG-grounded multi-hop reasoning on KG-grounded multi-hop reasoning 3-hop
60.32AccuracySPARK
Evaluation results
| Method | Links | |
|---|---|---|
| SPARKParams=4B2026.05 | 60.32 | |
| QWEN3-VL-4B + SPIN (NO KG)Params=4B2026.05 | 53.32 | |
| QWEN3-VL-4B + SFT ONLYParams=4B2026.05 | 53.25 | |
| QWEN3-VL-4B (VANILLA)Params=4B2026.05 | 51.8 |