Visual Perspective Taking on Perspective Taking
100Alignment ScoreLLaVA
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| LLaVAModel Scale=1.5-13B2026.01 | 100 | 0 | 50 | |
| Emb. Tokens (COCO)Protocol=Direct2026.01 | 100 | 100 | 100 | |
| Emb. Tokens (ViTPose)Protocol=CoT2026.01 | 100 | 100 | 100 | |
| Rotation TextProtocol=Direct2026.01 | 100 | 90 | 95 | |
| Rotation TokensProtocol=Direct2026.01 | 100 | 80 | 90 | |
| Rotation TokensProtocol=CoT2026.01 | 100 | 80 | 90 | |
| Emb. Tokens (COCO)Protocol=CoT2026.01 | 90 | 90 | 90 | |
| Emb. Tokens (ViTPose)Protocol=Direct2026.01 | 90 | 100 | 95 | |
| Rotation TextProtocol=CoT2026.01 | 90 | 70 | 80 | |
| Emb. TextProtocol=Direct2026.01 | 80 | 90 | 85 | |
| Emb. TextProtocol=CoT2026.01 | 80 | 80 | 80 |