List your company
ResearchFormerly WizWand

Mathematical Reasoning on Minerva (Accuracy @avg1)

54.9Accuracy (@avg1)TTSR
4.66817.70930.7543.791Sep 26, 2025Oct 22, 2025Nov 17, 2025Dec 14, 2025Jan 9, 2026Feb 4, 2026Mar 3, 2026
Updated 8mo ago

Evaluation results

MethodLinks
TTSRBackbone Model=Qwen3-8B-Base2026.0254.9
TTSRBackbone Model=Qwen3-4B-Base2026.0253
TTRLBackbone Model=Qwen3-8B-Base2026.0250.2
R-ZeroBackbone Model=Qwen3-8B-Base2026.0247.8
TTRLBackbone Model=Qwen3-4B-Base2026.0243.6
HACPOModel Backbone=Qwen3-8B-Base2026.0342.3
R-ZeroBackbone Model=Qwen3-4B-Base2026.0240.7
CAPO (Reinforce++)Backbone=Qwen2.5-7B-Math2025.1240.1
Base ModelBackbone Model=Qwen3-8B-Base2026.0239.6
GRPOBackbone=Qwen2.5-Math-7B2025.1138.6
KL-CovBackbone=Qwen2.5-Math-7B2025.1138.2
KL-CovBackbone=Qwen2.5-Base-7B2025.1137.9
CAPO (PPO)Backbone=Qwen2.5-7B-Math2025.1237.9
Forking TokensBackbone=Qwen2.5-Math-7B2025.1137.8
LESSBackbone=Qwen2.5-Math-7B2025.1137.8
Forking TokensBackbone=Qwen2.5-Base-7B2025.1137.5
GRPOBackbone=Qwen2.5-Base-7B2025.1137.1
LESSBackbone=Qwen2.5-Base-7B2025.1137.1
Reinforce++Backbone=Qwen2.5-7B-Math2025.1237.1
HACPOModel Backbone=Qwen3-4B-Base2026.0336.4
CAPO (RLOO)Backbone=Qwen2.5-7B-Math2025.1236
RLOOBackbone=Qwen2.5-7B-Math2025.1235.5
PPOBackbone=Qwen2.5-7B-Math2025.1234.2
CAPO (Reinforce++)Backbone=Qwen2.5-1.5B-Math2025.1234.2
TTSRBackbone Model=OctoThinker-8B-Hybrid-Base2026.0234.1
CAPO (RLOO)Backbone=Qwen2.5-1.5B-Math2025.1233.8
CAPO (GRPO)Backbone=Qwen2.5-7B-Math2025.1233.1
Base ModelBackbone Model=Qwen3-4B-Base2026.0232.4
CAPO (GRPO)Backbone=Qwen2.5-1.5B-Math2025.1232
Reinforce++Backbone=Qwen2.5-1.5B-Math2025.1232
LESSBackbone=Qwen2.5-Math-1.5B2025.1130.8
HACPOModel Backbone=Qwen3-1.7B-Base2026.0330.5
GRPOBackbone=Qwen2.5-7B-Math2025.1229.4
Forking TokensBackbone=Qwen2.5-Math-1.5B2025.1129
Base LLMBackbone=Qwen2.5-Math-7B2025.1128.7
RLOOBackbone=Qwen2.5-1.5B-Math2025.1228.7
KL-CovBackbone=Qwen2.5-Math-1.5B2025.1127.9
M_RLP + PostPre-training Condition=Reinforcement Learning from Pretraining, Post-training=SFT + RLVR2025.0927.8
GRPOBackbone=Qwen2.5-Math-1.5B2025.1127.2
GRPOBackbone=Qwen2.5-1.5B-Math2025.1226.8
TTRLBackbone Model=OctoThinker-8B-Hybrid-Base2026.0226.4
M_CPT + PostPre-training Condition=Continuous Pretraining, Post-training=SFT + RLVR2025.0926.1
CAPO (PPO)Backbone=Qwen2.5-1.5B-Math2025.1225.4
M_base + PostPre-training Condition=Base, Post-training=SFT + RLVR2025.0925.3
PPOBackbone=Qwen2.5-1.5B-Math2025.1224.6
CoTBackbone=Qwen2.5-1.5B-Math2025.1224.3
R-ZeroBackbone Model=OctoThinker-8B-Hybrid-Base2026.0222.9
CoTBackbone=Qwen2.5-7B-Math2025.1222.5
M_RLPPre-training Condition=Reinforcement Learning from Pretraining, Post-training=None2025.0921.19
HACPOModel Backbone=Llama3.2-3B-Instruct2026.0319.9
M_CPTPre-training Condition=Continuous Pretraining, Post-training=None2025.0919.03
Base LLMBackbone=Qwen2.5-Base-7B2025.1118
Base LLMBackbone=Qwen2.5-Math-1.5B2025.1115.8
M_basePre-training Condition=Base, Post-training=None2025.0915.3
Base ModelBackbone Model=OctoThinker-8B-Hybrid-Base2026.0215.2
HACPOModel Backbone=Llama3.2-1B-Instruct2026.038.1
HACPOModel Backbone=Llama3.2-1B-Instruct, Experimental Context=Heterogeneous Setup2026.036.6