Active Reasoning on ARBench-GN
45.3AccuracyTeamTR (Homogeneous)
Evaluation Results
| Method | Links | |
|---|---|---|
| TeamTR (Homogeneous)Size=3×8B2026.05 | 45.3 | |
| o4-miniSize=–2026.05 | 45.1 | |
| TeamTR (Heterogeneous)Size=1.7B+8B+14B2026.05 | 44.5 | |
| GPT-4o-miniSize=–2026.05 | 43.6 | |
| Debate + JudgeSize=3×8B2026.05 | 43.5 | |
| Role-playSize=3×8B2026.05 | 43.1 | |
| Qwen3-A3BSize=30B2026.05 | 42.3 | |
| DebateSize=3×8B2026.05 | 42.3 | |
| Qwen3 (thinking)Size=32B2026.05 | 41.1 | |
| DeepSeek-R1 DistillSize=70B2026.05 | 40.7 | |
| QwQSize=32B2026.05 | 39.5 | |
| DAPOSize=8B2026.05 | 39.3 | |
| GRPOSize=8B2026.05 | 38.9 |