List your company

Multi-Task Evaluation on Countdown, GSM8K, OlympiadBench, ROCStories, and MBPP (test)

56.3Overall ScoreRandOpt K=50
41.32445.21249.152.988May 29, 2026

Evaluation results

MethodLinks
RandOpt K=50Training cost (forward-pass-equivalent FLOPs)=1×10⁶, Inference cost (forward passes per test example)=502026.0556.3
CoRPTraining cost (forward-pass-equivalent FLOPs)=1×10⁵, Inference cost (forward passes per test example)=12026.0550
GRPOTraining cost (forward-pass-equivalent FLOPs)=1×10⁶, Inference cost (forward passes per test example)=12026.0548.8
PPOTraining cost (forward-pass-equivalent FLOPs)=1×10⁶, Inference cost (forward passes per test example)=12026.0547.7
RandOpt K=1Training cost (forward-pass-equivalent FLOPs)=1×10⁶, Inference cost (forward passes per test example)=12026.0543.5
BaseTraining cost (forward-pass-equivalent FLOPs)=–, Inference cost (forward passes per test example)=12026.0541.9