List your company

Mathematical Reasoning on Combined Mathematical Reasoning Benchmarks

75.81Average AccuracySRaR
19.015633.760348.50563.2497May 12, 2026May 13, 2026May 14, 2026May 15, 2026May 16, 2026May 17, 2026

Evaluation results

MethodLinks
SRaRModel Size=32B2026.0575.81
RaRModel Size=32B2026.0573.06
SRaRModel Size=8B2026.0572.53
DAPOModel Size=32B2026.0572.31
RaRModel Size=8B2026.0568.96
RGR-GRPOModel Size=32B2026.0567.73
DAPOModel Size=8B2026.0567.26
RGR-GRPOModel Size=8B2026.0563.24
GRPO-VPSModel Size=8B2026.0561.68
GRPOModel Size=32B2026.0561.65
GRPO-VPSModel Size=32B2026.0561.4
GRPOModel Size=8B2026.0558.54
BaselineModel Size=32B2026.0554.63
BaselineModel Size=8B2026.0553.4
OPEFOBackbone=Qwen2.5-Math-7B2026.0552.4
OPEFOBackbone=Qwen3-Base-4B2026.0551.9
Clip–CovBackbone=Qwen2.5-Math-7B2026.0550.7
KL–CovBackbone=Qwen2.5-Math-7B2026.0550.5
GRPO (Strict on-policy)Backbone=Qwen2.5-Math-7B2026.0550.1
GRPO (Strict on-policy)Backbone=Qwen3-Base-4B2026.0550.1
Clip-higherBackbone=Qwen2.5-Math-7B2026.0549.8
KL–CovBackbone=Qwen3-Base-4B2026.0549.2
Clip-higherBackbone=Qwen3-Base-4B2026.0549.1
Clip–CovBackbone=Qwen3-Base-4B2026.0548.9
Entropy-RegBackbone=Qwen3-Base-4B2026.0548
Entropy-RegBackbone=Qwen2.5-Math-7B2026.0547.8
GRPOBackbone=Qwen2.5-Math-7B2026.0545.5
GRPOBackbone=Qwen3-Base-4B2026.0544.8
Qwen3-Base-4BBackbone=Qwen3-Base-4B2026.0536.4
Qwen2.5-Math-7BBackbone=Qwen2.5-Math-7B2026.0521.2