Mathematical Reasoning on MATH-500 (Consensus & ARBITER Evaluation)
73Accuracy (Raw Consensus)Raw consensus
Evaluation results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| Raw consensusModel=Phi-4, Policy=no-framing majority baseline, Sample count (N)=5002026.05 | 73 | — | — | — | — | — | — | |
| ARBITER-∆Model=Phi-4, Policy=fixed framed+guided parameter-free, Sample count (N)=5002026.05 | 73 | 73.2 | 0.2 | 22 | 5 | 4 | 1 | |
| Raw consensusModel=Qwen3-4B, Policy=no-framing majority baseline, Sample count (N)=5002026.05 | 69.2 | — | — | — | — | — | — | |
| ARBITER-∆Model=Qwen3-4B, Policy=fixed framed+guided parameter-free, Sample count (N)=5002026.05 | 69.2 | 69.2 | 0 | 11 | 4 | 4 | 0 | |
| Raw consensusModel=Llama-3.1-8B, Policy=no-framing majority baseline, Sample count (N)=5002026.05 | 51.6 | — | — | — | — | — | — | |
| ARBITER-∆Model=Llama-3.1-8B, Policy=fixed framed+guided parameter-free, Sample count (N)=5002026.05 | 51.6 | 54.6 | 3 | 44 | 19 | 4 | 15 |