Active Reasoning on ARBench SP
43.1Accuracyo4-mini
Evaluation Results
| Method | Links | |
|---|---|---|
| o4-miniSize=–2026.05 | 43.1 | |
| TeamTR (Homogeneous)Size=3×8B2026.05 | 43.1 | |
| TeamTR (Heterogeneous)Size=1.7B+8B+14B2026.05 | 42.3 | |
| GPT-4o-miniSize=–2026.05 | 40.8 | |
| Debate + JudgeSize=3×8B2026.05 | 39.4 | |
| Qwen3 (thinking)Size=32B2026.05 | 38.9 | |
| Role-playSize=3×8B2026.05 | 38.7 | |
| Qwen3-A3BSize=30B2026.05 | 38.4 | |
| DebateSize=3×8B2026.05 | 38 | |
| QwQSize=32B2026.05 | 36.1 | |
| DAPOSize=8B2026.05 | 35.9 | |
| DeepSeek-R1 DistillSize=70B2026.05 | 35.4 | |
| GRPOSize=8B2026.05 | 35.2 |