Mathematical Reasoning on GSM8K (Acc(%), Avg.Steps(n))
98.5Accuracy (%)Univ. Self-Consistency
Evaluation results
| Method | Links | ||
|---|---|---|---|
| Univ. Self-ConsistencyBackbone=o4-mini2026.04 | 98.5 | 5.7 | |
| Best-of-NBackbone=o4-mini2026.04 | 98 | 5.8 | |
| CRAFTBackbone=o4-mini2026.04 | 98 | 5.5 | |
| Self-AggregationBackbone=o4-mini2026.04 | 96.6 | 8.2 | |
| CRAFTBackbone=GPT-5.4-nano2026.04 | 96 | 5.4 | |
| RAPBackbone=GPT-5.4-nano2026.04 | 95.8 | 8.9 | |
| Self-ConsistencyBackbone=o4-mini2026.04 | 95.6 | 5.3 | |
| RAPBackbone=o4-mini2026.04 | 95 | 5.9 | |
| Best-of-NBackbone=GPT-5.4-nano2026.04 | 94.8 | 19.4 | |
| Self-RefineBackbone=o4-mini2026.04 | 93.2 | 5.8 | |
| Faithful CoTBackbone=o4-mini2026.04 | 93.2 | 4.4 | |
| Self-ConsistencyBackbone=GPT-5.4-nano2026.04 | 93 | 17.9 | |
| Self-AggregationBackbone=GPT-5.4-nano2026.04 | 93 | 11.3 | |
| Univ. Self-ConsistencyBackbone=GPT-5.4-nano2026.04 | 92 | 17.6 | |
| Tree-of-ThoughtBackbone=o4-mini2026.04 | 91.8 | 6.5 | |
| Self-RefineBackbone=GPT-5.4-nano2026.04 | 91.4 | 19.1 | |
| Tree-of-ThoughtBackbone=GPT-5.4-nano2026.04 | 90.6 | 6.2 | |
| Faithful CoTBackbone=GPT-5.4-nano2026.04 | 90.2 | 7 | |
| Self-Eval Beam SearchBackbone=GPT-5.4-nano2026.04 | 79 | 4.4 | |
| Self-Eval Beam SearchBackbone=o4-mini2026.04 | 37.8 | 2.2 |