Preference-Aware Planning on 10,000 City Samples (test)
95.3ConnectivityQwen3-4B-Joint
Evaluation results
| Method | Links | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| Qwen3-4B-JointModel scale=4B, Fine-tuning strategy=Joint SFT on combined data of all three tasks2026.05 | 95.3 | 97.4 | 88.5 | 0.709 | 0.726 | 52.6 | 92.8 | 1.88 | 90.5 | |
| Qwen3-4BModel scale=4B2026.05 | 93.2 | 96.5 | 84.6 | 0.705 | 0.716 | 50.4 | 92.5 | 2.05 | 89.8 | |
| Qwen3-4B-25Model scale=4B, Continuous Pre-Training (CPT) data percentage=25% session data2026.05 | 90.1 | 94.2 | 77.4 | 0.683 | 0.683 | 43.9 | 90.9 | 2.49 | 88.8 | |
| Qwen3-1.7BModel scale=1.7B2026.05 | 88.5 | 92.2 | 75.8 | 0.656 | 0.651 | 39.4 | 91.6 | 2.29 | 81.6 | |
| Qwen3-0.6BModel scale=0.6B2026.05 | 85 | 92.2 | 73.3 | 0.65 | 0.643 | 38.7 | 90.8 | 2.74 | 81.8 |