List your company

Mathematical Reasoning on AIME 24 (Accuracy@4, GRPO Speedup)

18.3Accuracy (mean@4)DAPO
2.76.7510.814.85May 8, 2026

Evaluation results

MethodLinks
DAPOBackbone=Qwen3-4B-base2026.0518.30.63
VIPBackbone=Qwen3-4B-base2026.0516.70.75
DUETBackbone=Qwen3-4B-base, Budget=Full2026.0514.21.43
GRPOBackbone=Qwen3-4B-base2026.0513.31
DUETBackbone=Qwen3-4B-base, Budget=50%2026.0513.32.38
DUETBackbone=Qwen3-1.7B-base, Budget=Full2026.0512.51.62
ARRoLBackbone=Qwen3-4B-base2026.0511.71.19
DUETBackbone=Qwen3-1.7B-base, Budget=50%2026.0510.82.51
BaselineBackbone=Qwen3-4B-base2026.0510—
VIPBackbone=Llama-3.2-3B-Instruct2026.059.20.95
ARRoLBackbone=Qwen3-1.7B-base2026.058.31.3
DAPOBackbone=Qwen3-1.7B-base2026.057.50.81
GRPOBackbone=Llama-3.2-3B-Instruct2026.057.51
DAPOBackbone=Llama-3.2-3B-Instruct2026.057.50.82
DUETBackbone=Llama-3.2-3B-Instruct, Budget=Full2026.057.51.26
GRPOBackbone=Qwen3-1.7B-base2026.056.71
ARRoLBackbone=Llama-3.2-3B-Instruct2026.056.71.13
DUETBackbone=Llama-3.2-3B-Instruct, Budget=50%2026.056.72.04
BaselineBackbone=Qwen3-1.7B-base2026.055.8—
VIPBackbone=Qwen3-1.7B-base2026.055.80.95
BaselineBackbone=Llama-3.2-3B-Instruct2026.053.3—