Long-horizon task execution on LongAct Bench (detailed split)
59GC (Avg)GPT-5
Evaluation Results
| Method | Links | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| GPT-5Architecture=Agentic (HoloMind)2026.05 | 59 | 66.7 | 82.9 | 75.2 | 60 | 16 | 1,982 | 25.3 | 1.7 | |
| Qwen3-VL-32BArchitecture=Agentic (HoloMind)2026.05 | 51.2 | 52.7 | 71.2 | 73.4 | 25 | 15 | 1,692 | 28.7 | 1.61 | |
| GPT-5-miniArchitecture=Agentic (HoloMind)2026.05 | 38.4 | 40.5 | 58.9 | 66.7 | 25 | 9 | 2,304 | 21.5 | 0.81 | |
| Qwen3-VL-8BArchitecture=Agentic (HoloMind)2026.05 | 24.5 | 27.4 | 55 | 48.7 | 31.3 | 3 | 1,896 | 30.6 | 0.99 | |
| Qwen3-VL-2BArchitecture=Agentic (HoloMind)2026.05 | 7.16 | 7.04 | 8.77 | 37.8 | 0 | 0 | 844 | 4.81 | 0.59 | |
| Qwen3-VL-32BArchitecture=Pure VLM, Agentic Architecture=None2026.05 | 6.14 | 3.54 | 18.8 | 7.32 | 0 | 0 | 2,981 | 3.1 | -0.32 | |
| Qwen3-VL-8BArchitecture=Pure VLM, Agentic Architecture=None2026.05 | 0.74 | 0.15 | 6.93 | 0 | 0 | 0 | 2,598 | 0.26 | -0.08 |