Compositional Reasoning on NaturalBench
35.5AccuracyInternVL3.5-14B +FINER-Tuning
Evaluation results
| Method | Links | |
|---|---|---|
| InternVL3.5-14B +FINER-TuningBackbone=InternVL3.5-14B, Training Strategy=FINER-Tuning2026.03 | 35.5 | |
| Qwen2.5-VL-7BBackbone=Qwen2.5-VL-7B2026.03 | 34.1 | |
| Qwen2.5-VL-7B +FINER-TuningBackbone=Qwen2.5-VL-7B, Training Strategy=FINER-Tuning2026.03 | 34.1 | |
| InternVL3.5-8B +FINER-TuningBackbone=InternVL3.5-8B, Training Strategy=FINER-Tuning2026.03 | 31.1 | |
| InternVL3.5-14BBackbone=InternVL3.5-14B2026.03 | 30.7 | |
| InternVL3.5-8BBackbone=InternVL3.5-8B2026.03 | 30.4 | |
| OmniLMM-12BBackbone=OmniLMM-12B2026.03 | 26.9 | |
| LLaVA-1.6-7B +FINER-TuningBackbone=LLaVA-1.6-7B, Training Strategy=FINER-Tuning2026.03 | 19.8 | |
| OmniLMM-12B +RLAIF-VBackbone=OmniLMM-12B, Training Strategy=RLAIF-V2026.03 | 19.4 | |
| LLaVA-1.6-7BBackbone=LLaVA-1.6-7B2026.03 | 15.7 |