List your company

Mathematical Reasoning on Minerva (Mean@32 accuracy)

36.5Mean@32 AccuracyBAPO
27.24429.64732.0534.453May 8, 2026

Evaluation results

MethodLinks
BAPOBackbone=Qwen3-8B-Instruct2026.0536.5
HTPOBackbone=Qwen3-8B-Instruct2026.0536.5
GSPOBackbone=Qwen3-8B-Instruct2026.0535.9
GRPO†Backbone=Qwen3-8B-Instruct2026.0534.8
80/20-RuleBackbone=Qwen3-8B-Instruct2026.0534.3
DAPOBackbone=Qwen3-8B-Instruct2026.0533.8
SAPOBackbone=Qwen3-8B-Instruct2026.0533.1
GSPOBackbone=Qwen3-8B-Base2026.0530.3
HTPOBackbone=Qwen3-8B-Base2026.0530.3
SAPOBackbone=Qwen3-8B-Base2026.0530.1
DAPOBackbone=Qwen3-8B-Base2026.0529.4
GRPO†Backbone=Qwen3-8B-Base2026.0528.5
BAPOBackbone=Qwen3-8B-Base2026.0528.3
80/20-RuleBackbone=Qwen3-8B-Base2026.0527.6