Goal reconstruction on OneStop
54.8Accuracy (Question Word)DalEye-LLaVA
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| DalEye-LLaVAInput Modality=Eye-tracking + Text2025.05 | 54.8 | 46.6 | 23.1 | |
| DalEye-LlamaInput Modality=Eye-tracking + Text2025.05 | 54.5 | 45 | 25.3 | |
| Text-only Llama 3.1Input Modality=Text-only2025.05 | 54 | 45.4 | 23.6 | |
| Text-only LLaVA OneVisionInput Modality=Text-only2025.05 | 49.9 | 42.4 | 21 | |
| DalEye-GPTInput Modality=Eye-tracking + Text2025.05 | 49.5 | 48.8 | 25.7 | |
| Text-only GPT-4o-miniInput Modality=Text-only2025.05 | 45 | 43 | 24 | |
| Incorrect Human (different critical span)Input Modality=Human Performance2025.05 | 40.7 | 41 | 21.9 | |
| Gemini few-shotPrompting Strategy=few-shot2025.05 | 39.5 | 41.8 | 27.2 | |
| Gemini zero-shotPrompting Strategy=zero-shot2025.05 | 37.8 | 33.5 | 24.7 | |
| Incorrect Human (same critical span)Input Modality=Human Performance2025.05 | 37 | 34.9 | 29.8 | |
| Arbitrary Gemini 3Prompting Strategy=Arbitrary2025.05 | 32.7 | 36.7 | 19.4 |