Skill Evaluation on LangR
0.8659Object Perception (Grounding) AccuracyBC (InternVL3-8B)
Evaluation results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| BC (InternVL3-8B)Backbone=InternVL3-8B2026.04 | 0.8659 | 0.8893 | 1 | 0.3878 | 0.1414 | 0.7367 | 0.9804 | |
| Ours (InternVL3-8B)Backbone=InternVL3-8B2026.04 | 0.865 | 0.932 | 1 | 1 | 0.4984 | 0.7407 | 0.9927 | |
| Gemini-2.5-ProModel Class=Closed2026.04 | 0.4568 | 0.5977 | 0.954 | 0.8214 | 0.3035 | 0.388 | 0.9707 | |
| Qwen3-VL-8BModel Class=Open2026.04 | 0.3841 | 0.2994 | 0.91 | 0.6327 | 0.2722 | 0.606 | 0.7213 | |
| GPT-o3Model Class=Closed2026.04 | 0.3517 | 0.2077 | 0.958 | 0.7296 | 0.273 | 0.5967 | 0.8557 | |
| GPT-5Model Class=Closed2026.04 | 0.3507 | 0.1235 | 0.874 | 0.625 | 0.2829 | 0.6053 | 0.8802 | |
| Gemini-2.5-FlashModel Class=Closed2026.04 | 0.3043 | 0.2408 | 0.944 | 0.7194 | 0.2755 | 0.592 | 0.7726 | |
| InternVL3-8BModel Class=Open2026.04 | 0.2903 | 0.0862 | 0.746 | 0.48 | 0.131 | 0.529 | 0.6968 | |
| Ours* (ALFRED-tuned)Model Class=Cross, Fine-tuning=ALFRED-tuned2026.04 | 0.2681 | 0.5835 | 0.866 | 0.7781 | 0.4752 | — | 0.7706 | |
| InternVL2.5-8BModel Class=Open2026.04 | 0.2599 | 0.0112 | 0.75 | 0.44 | 0.1259 | 0.4753 | 0.3032 | |
| Claude-4.5-SonnetModel Class=Closed2026.04 | 0.2496 | 0 | 0.644 | 0.3827 | 0.215 | 0.438 | 0.89 | |
| Qwen2.5-VL-7BModel Class=Open2026.04 | 0.2164 | 0 | 0.64 | 0.3622 | 0.2204 | 0.5707 | 0.7946 | |
| LLaVA-OneVision-7BModel Class=Open2026.04 | 0.1115 | 0.0852 | 0.53 | 0.426 | 0.0019 | 0.3953 | 0.4034 |