Mathematical Reasoning on GSM8K (Avg@16 Accuracy & ATC)
0.928Avg@16 AccuracyAdaTIR-7B
Evaluation results
| Method | Links | ||
|---|---|---|---|
| AdaTIR-7BBackbone=Qwen2.5-7B, Training Strategy=AdaTIR2026.01 | 0.928 | 0.02 | |
| Qwen2.5-7B-Instruct (GRPO)Backbone=Qwen2.5-7B, Training Strategy=GRPO2026.01 | 0.91 | 0.83 | |
| AdaTIR-3BBackbone=Qwen2.5-3B, Training Strategy=AdaTIR2026.01 | 0.871 | 0.22 | |
| Qwen2.5-3B-Instruct (GRPO)Backbone=Qwen2.5-3B, Training Strategy=GRPO2026.01 | 0.865 | 1.19 | |
| Qwen2.5-7B-InstructBackbone=Qwen2.5-7B, Training Strategy=Instruct2026.01 | 0.619 | — | |
| Qwen2.5-3B-InstructBackbone=Qwen2.5-3B, Training Strategy=Instruct2026.01 | 0.5 | — | |
| Qwen2.5-3B-Instruct-TIRBackbone=Qwen2.5-3B, Training Strategy=TIR2026.01 | 0.36 | 1.58 | |
| Qwen2.5-7B-Instruct-TIRBackbone=Qwen2.5-7B, Training Strategy=TIR2026.01 | 0.173 | 0.51 |