Visual Reasoning on Sokoban
85.71AccuracyGemini-1.5-Flash
Evaluation Results
| Method | Links | |
|---|---|---|
| Gemini-1.5-FlashModel Category=Frontier Models2026.06 | 85.71 | |
| Qwen2.5-7BModel Category=Open-Source Models2026.06 | 80.95 | |
| Seed-1.5-27BModel Category=Frontier Models2026.06 | 69.05 | |
| Gemma-2-9BModel Category=Open-Source Models2026.06 | 66.67 | |
| Ours (Stage 2: optimize Lund→gen)Optimization Stage=Stage 2 (Flow-GRPO), Model Category=Our Models2026.06 | 50 | |
| Ours (Stage 1: optimize Lgen→und)Optimization Stage=Stage 1 (Reflective SFT), Model Category=Our Models2026.06 | 43.65 | |
| Gemini-1.5-Flash-8BModel Category=Frontier Models2026.06 | 37.3 | |
| ThinkMorphModel Category=Open-Source Models, Interleave Thinking=false2026.06 | 22.22 | |
| ThinkMorphModel Category=Open-Source Models, Interleave Thinking=true2026.06 | 18.25 | |
| Bagel-7B-MoTModel Category=Open-Source Models2026.06 | 16.67 |