Math Problem Solving on GSM8K (Accuracy)
91.5AccuracyTaH+
Evaluation results
| Method | Links | ||
|---|---|---|---|
| TaH+Param.=4B2025.11 | 91.5 | — | |
| SoftThinkParam.=4B2025.11 | 90.9 | — | |
| StandardParam.=4B2025.11 | 90.5 | — | |
| TaHParam.=4B2025.11 | 90.4 | — | |
| TaH+Param.=1.7B2025.11 | 85.8 | — | |
| RoutingParam.=4B2025.11 | 85.3 | — | |
| TaHParam.=1.7B2025.11 | 84.5 | — | |
| StandardParam.=1.7B2025.11 | 82.1 | — | |
| SoftThinkParam.=1.7B2025.11 | 79.6 | — | |
| AlwaysThinkParam.=1.7B2025.11 | 79.3 | — | |
| RoutingParam.=1.7B2025.11 | 71.2 | — | |
| TaH+Param.=0.6B2025.11 | 68.8 | — | |
| TaHParam.=0.6B2025.11 | 64.4 | — | |
| StandardParam.=0.6B2025.11 | 62.5 | — | |
| SoftThinkParam.=0.6B2025.11 | 61.3 | — | |
| AlwaysThinkParam.=0.6B2025.11 | 54.6 | — | |
| RoutingParam.=0.6B2025.11 | 45.6 | — | |
| deepseek-moe-16b-chatParameters=16b, Architecture=moe, Type=chat2026.05 | — | 51.5 | |
| EngGPT2-16B-A3BParameters=16B, Architecture=A3B2026.05 | — | 86.3 | |
| FastwebMIIA-7BParameters=7B2026.05 | — | 53.1 | |
| gemma-3-12b-itParameters=12b, Type=it2026.05 | — | 90.7 | |
| gemma-3-4b-itParameters=4b, Type=it2026.05 | — | 81.6 | |
| GPT-5 nanoSize=nano2026.05 | — | 94.4 | |
| gpt-oss-20bParameters=20b2026.05 | — | 92.5 | |
| Llama-3.1-8B-InstructParameters=8B, Type=Instruct2026.05 | — | 85.7 | |
| Llama-3.2-3B-InstructParameters=3B, Type=Instruct2026.05 | — | 77.9 | |
| LLaMAntino-3-ANITA-8B-Inst-DPO-ITAParameters=8B, Type=Inst, Optimization=DPO, Language=ITA2026.05 | — | 69 | |
| Minerva-7B-instruct-v1.0Parameters=7B, Type=instruct, Version=v1.02026.05 | — | 9 | |
| Ministral-3-8B-Instruct-2512-BF16Parameters=8B, Type=Instruct, Precision=BF162026.05 | — | 75.1 | |
| Moonlight-16B-A3B-InstructParameters=16B, Architecture=A3B, Type=Instruct2026.05 | — | 73.3 | |
| Qwen3-4BParameters=4B2026.05 | — | 86.8 | |
| Qwen3-8BParameters=8B2026.05 | — | 94.3 | |
| Velvet-14BParameters=14B2026.05 | — | 61 |