List your company

Mathematical Reasoning on MATH 500 (Accuracy@4 and Speedup)

78.2Accuracy (mean@4)DUET
35.45646.55357.6568.747May 8, 2026

Evaluation results

MethodLinks
DUETBackbone=Qwen3-4B-base, Budget=Full2026.0578.21.43
DUETBackbone=Qwen3-4B-base, Budget=50%2026.0576.72.38
GRPOBackbone=Qwen3-4B-base2026.0574.51
VIPBackbone=Qwen3-4B-base2026.0574.40.75
ARRoLBackbone=Qwen3-4B-base2026.0573.91.19
DAPOBackbone=Qwen3-4B-base2026.0573.20.63
DUETBackbone=Qwen3-1.7B-base, Budget=50%2026.0562.12.51
DUETBackbone=Qwen3-1.7B-base, Budget=Full2026.0561.61.62
VIPBackbone=Qwen3-1.7B-base2026.0558.60.95
ARRoLBackbone=Qwen3-1.7B-base2026.0558.41.3
DAPOBackbone=Qwen3-1.7B-base2026.0557.60.81
GRPOBackbone=Qwen3-1.7B-base2026.0556.81
BaselineBackbone=Qwen3-4B-base2026.0554.3—
DUETBackbone=Llama-3.2-3B-Instruct, Budget=Full2026.0545.61.26
ARRoLBackbone=Llama-3.2-3B-Instruct2026.0545.31.13
BaselineBackbone=Qwen3-1.7B-base2026.0545.1—
DUETBackbone=Llama-3.2-3B-Instruct, Budget=50%2026.0544.52.04
GRPOBackbone=Llama-3.2-3B-Instruct2026.0544.41
DAPOBackbone=Llama-3.2-3B-Instruct2026.0544.10.82
VIPBackbone=Llama-3.2-3B-Instruct2026.0543.60.95
BaselineBackbone=Llama-3.2-3B-Instruct2026.0537.1—