ResearchBenchmarksDownstream accuracy extrapolation on OpenBookQAFollow0RMSECompute-0.003440.019780.0430.06622May 18, 2026Evaluation ResultsMethodMethodLinksRMSEComputePredictor=ComputePredictor=Compute2026.050ProxyPredictor=ProxyPredictor=Proxy2026.050.069CE LossPredictor=CE LossPredictor=CE Loss2026.050.086