Mathematical Reasoning on GSM8K (Acc %, Tokens)
78.8AccuracyTL-1.5B-RL
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| TL-1.5B-RLBudget=100%2025.10 | 78.8 | 573 | |
| TL-Warmup + DiffAdaptBudget=100%2025.10 | 78.6 | 364 | |
| TL-Warmup + DiffAdaptBudget=72.2%2025.10 | 78.2 | 361 | |
| TL-Warmup + DiffAdaptBudget=50.0%2025.10 | 77 | 357 | |
| TL-1.5B-RLBudget=72.2%2025.10 | 74.3 | 510 | |
| TL-1.5B-RLBudget=50.0%2025.10 | 72 | 465 | |
| TL-Warmup + DiffAdaptBudget=33.3%2025.10 | 67 | 336 | |
| TL-1.5B-RLBudget=33.3%2025.10 | 66.1 | 409 |