List your company
ResearchFormerly WizWand

Mathematical Reasoning on MATH500 (val)

87.64AccuracyRLOO
38.302451.111263.9276.7288Nov 28, 2025Dec 20, 2025Jan 12, 2026Feb 3, 2026Feb 26, 2026Mar 20, 2026Apr 12, 2026
Updated 6mo ago

Evaluation results

MethodLinks
RLOOAlgorithm=RLOO, Base Model=Qwen3-8B-Base, Sampled generations=16, RL Setup=Zero RL2026.0487.64
VC-PPOAlgorithm=VC-PPO, Base Model=Qwen3-8B-Base, Sampled generations=16, RL Setup=Zero RL2026.0487.54
GenACAlgorithm=GenAC, Base Model=Qwen3-8B-Base, Sampled generations=16, RL Setup=Zero RL2026.0487.48
GRPOAlgorithm=GRPO, Base Model=Qwen3-8B-Base, Sampled generations=16, RL Setup=Zero RL2026.0487.2
SyncLag (k)=0, Steps=400, GPU hours=134.42026.0272
VCPOLag (k)=10, Steps=400, GPU hours=92.82026.0271.6
OBLR-POModel=Qwen3-8B-Base2025.1170.4
ReMaxModel=Qwen3-8B-Base2025.1169.8
GRPOModel=Qwen3-8B-Base2025.1169.6
RLOOModel=Qwen3-8B-Base2025.1168.8
RLOOModel=Qwen3-4B-Base2025.1167.8
OBLR-POModel=Qwen3-4B-Base2025.1167.8
GRPOModel=Qwen3-4B-Base2025.1167.6
ReMaxModel=Qwen3-4B-Base2025.1165.6
PPOModel=Qwen3-8B-Base2025.1164.6
PPOModel=Qwen3-4B-Base2025.1159.2
Base2026.0240.2