Compositional Reasoning on Winoground 400 pairs (test)
32.25Text ScoreLLaVA-1.5-7B + Eucl. visual relations
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| LLaVA-1.5-7B + Eucl. visual relationsScoring protocol=sequence log-likelihood scoring, Geometry=Euclidean, Prefix tokens=true2026.06 | 32.25 | 24.25 | 16.25 | |
| HyperVisScoring protocol=sequence log-likelihood scoring, Geometry=Hyperbolic, Prefix tokens=true2026.06 | 31.75 | 30.75 | 18.5 | |
| LLaVA-1.5-7BScoring protocol=sequence log-likelihood scoring, Configuration=baseline2026.06 | 31.5 | 32.5 | 19 | |
| LLaVA-1.5-7B + LoRA onlyScoring protocol=sequence log-likelihood scoring, Relational loss=none2026.06 | 30.5 | 32.5 | 19.5 |