ResearchBenchmarksMathematical Reasoning on Countdown (val)Mathematical Reasoning on Countdown (val)Follow41.9AccuracyVCPO-0.01210.86921.7532.631Feb 19, 2026Evaluation resultsMethodMethodLinksAccuracyVCPOLag (k)=10, Steps=400,...Lag (k)=10, Steps=400, GPU hours=89.62026.0241.9SyncLag (k)=0, Steps=400,...Lag (k)=0, Steps=400, GPU hours=143.22026.0238.4Base2026.021.6