List your company
ResearchFormerly WizWand

Math Reasoning on AIME 24 (Accuracy)

83.3AccuracyBaseline
-3.33219.15941.6564.141May 16, 2025Jul 20, 2025Sep 24, 2025Nov 29, 2025Feb 3, 2026Apr 10, 2026Jun 15, 2026
Updated 3mo ago

Evaluation results

MethodLinks
BaselineSparsity=0%2025.1183.3
PuzzleMoESparsity=25%2025.1171.1
VRPOModel Backbone=Qwen3-8B(Strong Model), Optimization Method=VRPO2025.0846.67
Reinforce++Model Backbone=Qwen3-8B(Strong Model), Optimization Method=Reinforce++2025.0845
Dr.GRPOModel Backbone=Qwen3-8B(Strong Model), Optimization Method=Dr.GRPO2025.0845
BaseModel Backbone=Qwen3-8B(Strong Model), Optimization Method=Base2025.0841.67
PPOModel Backbone=Qwen3-8B(Strong Model), Optimization Method=PPO2025.0836.67
GRPOModel Backbone=Qwen3-8B(Strong Model), Optimization Method=GRPO2025.0835
EGLR-OracleLModel=Qwen2.5-Math-7B-Instruct2026.0630
EGLR-OracleLModel=Qwen2.5-14B-Instruct2026.0626.7
VRPOModel Backbone=Qwen2.5-7B-Cold Start(Weak Model), Optimization Method=VRPO2025.0823.33
EGLR-OracleLModel=Qwen2.5-7B-Instruct2026.0623.3
EGLR-OracleLModel=Llama-3.1-8B-Instruct2026.0623.3
CFTTeacher Model=S1.1-32B2025.0520
CGDTeacher Model=Claude Sonnet 3.72025.0520
EGLR-OracleLModel=Qwen2.5-3B-Instruct2026.0620
EGLR (best L)Model=Qwen2.5-14B-Instruct2026.0620
EGLR-SCL (B = 10)Model=Qwen2.5-14B-Instruct2026.0620
EGLR-OracleLModel=Qwen2.5-Math-1.5B-Instruct2026.0620
EGLR (best L)Model=Qwen2.5-Math-7B-Instruct2026.0620
CFTTeacher Model=GPT-4o2025.0516.7
EGLR-SCL (B = 10)Model=Qwen2.5-3B-Instruct2026.0616.7
EGLR (best L)Model=Qwen2.5-7B-Instruct2026.0616.7
EGLR-SCL (B = 10)Model=Qwen2.5-7B-Instruct2026.0616.7
PPOModel Backbone=Qwen2.5-7B-Cold Start(Weak Model), Optimization Method=PPO2025.0813.33
Dr.GRPOModel Backbone=Qwen2.5-7B-Cold Start(Weak Model), Optimization Method=Dr.GRPO2025.0813.33
CGDTeacher Model=S1.1-32B2025.0513.3
EGLR (best L)Model=Qwen2.5-3B-Instruct2026.0613.3
GreedyModel=Qwen2.5-Math-1.5B-Instruct2026.0613.3
EGLR (best L)Model=Qwen2.5-Math-1.5B-Instruct2026.0613.3
GreedyModel=Qwen2.5-Math-7B-Instruct2026.0613.3
EGLR-SCL (B = 10)Model=Qwen2.5-Math-7B-Instruct2026.0613.3
EGLR (best L)Model=Llama-3.1-8B-Instruct2026.0613.3
Qwen2.5-Math-7B (Base)2025.0510
GreedyModel=Qwen2.5-14B-Instruct2026.0610
EGLR-SCL (B = 10)Model=Qwen2.5-Math-1.5B-Instruct2026.0610
EGLR-SCL (B = 10)Model=Llama-3.1-8B-Instruct2026.0610
GreedyModel=Qwen2.5-3B-Instruct2026.066.7
GreedyModel=Qwen2.5-7B-Instruct2026.066.7
BaseModel Backbone=Qwen2.5-7B-Cold Start(Weak Model), Optimization Method=Base2025.086.67
GRPOModel Backbone=Qwen2.5-7B-Cold Start(Weak Model), Optimization Method=GRPO2025.086.67
Reinforce++Model Backbone=Qwen2.5-7B-Cold Start(Weak Model), Optimization Method=Reinforce++2025.086.67
EGLR (best L)Model=Qwen2.5-0.5B-Instruct2026.063.3
EGLR-SCL (B = 10)Model=Qwen2.5-0.5B-Instruct2026.063.3
EGLR-OracleLModel=Qwen2.5-0.5B-Instruct2026.063.3
GreedyModel=Mistral-7B-Instruct-v0.32026.063.3
EGLR (best L)Model=Mistral-7B-Instruct-v0.32026.063.3
EGLR-SCL (B = 10)Model=Mistral-7B-Instruct-v0.32026.063.3
EGLR-OracleLModel=Mistral-7B-Instruct-v0.32026.063.3
GreedyModel=Qwen2.5-0.5B-Instruct2026.060
GreedyModel=Llama-3.1-8B-Instruct2026.060
HC-SMoESparsity=25%2025.110