Image-to-Sequence reconstruction on ABC out-of-distribution set
66.2Success Rate (%)GPT-5.2 High
Evaluation Results
| Method | Links | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| GPT-5.2 HighReasoning level=High2026.04 | 66.2 | 34.4 | 28.9 | 50.9 | 73 | 0.22 | 0.03 | 0.12 | 0.33 | |
| GPT-5.2 MediumReasoning level=Medium2026.04 | 62.6 | 34.6 | 28.5 | 55.7 | 74.7 | 0.22 | 0.03 | 0.13 | 0.34 | |
| Qwen3-VL-2B (fine-tuned)Fine-tuning status=fine-tuned2026.04 | 61 | 37.7 | 30.3 | 64.4 | 85.4 | 0.1 | 0.032 | 0.11 | 0.29 | |
| Qwen3-VL-2B (base)Fine-tuning status=base2026.04 | 5.4 | 13.1 | 7.3 | 18.5 | 35.5 | 0.45 | 0.35 | 0.65 | 1.01 |