List your company
ResearchFormerly WizWand

Mathematical Reasoning on LMB (LiveMathBench)

35.8AccuracyRePro
6.36814.00921.6529.291Dec 1, 2025
Updated 8mo ago

Evaluation results

MethodLinks
ReProBackbone=Qwen3-8B, Base Training=GRPO2025.1235.8
OriginalBackbone=Qwen3-8B2025.1235.3
GRPOBackbone=Qwen3-8B2025.1235.1
RF++ B + ReProBackbone=Qwen3-1.7B2025.1222
PPOBackbone=Qwen3-1.7B2025.1220.5
PPO + ReProBackbone=Hunyuan-1.8B-Instruct2025.1220.5
GRPO + ReProBackbone=Qwen3-1.7B2025.1219.5
PPOBackbone=Hunyuan-1.8B-Instruct2025.1219.5
OriginalBackbone=Qwen3-1.7B2025.1218.8
PPO + ReProBackbone=Qwen3-1.7B2025.1218.8
GRPOBackbone=Qwen3-1.7B2025.1218.8
RF++ BBackbone=Qwen3-1.7B2025.1218.5
RF++ B + ReProBackbone=Hunyuan-1.8B-Instruct2025.1217.8
RF++ BBackbone=Hunyuan-1.8B-Instruct2025.1217.5
GRPO + ReProBackbone=Hunyuan-1.8B-Instruct2025.1217.5
GRPOBackbone=Hunyuan-1.8B-Instruct2025.1216
PPO + ReProBackbone=MobileLLM-R1-950M2025.1215.8
OriginalBackbone=Hunyuan-1.8B-Instruct2025.1215
ReProBackbone=Qwen3-4B-Base, Sample Count=Avg@4, Evaluation Domain=In-domain, Base Method=GRPO2025.1214.5
PPOBackbone=MobileLLM-R1-950M2025.1214.5
GRPOBackbone=Qwen3-4B-Base, Sample Count=Avg@4, Evaluation Domain=In-domain2025.1212.5
OriginalBackbone=MobileLLM-R1-950M2025.1210.5
OriginalBackbone=Qwen3-4B-Base, Sample Count=Avg@4, Evaluation Domain=In-domain2025.127.5