ResearchBenchmarksMath Reasoning on MATH-500 (held-out)Follow4.79Accuracy Delta (pp)alpha0.03721.27112.5053.7389May 14, 2026Evaluation ResultsMethodMethodLinksAccuracy Delta (pp)95% CIp-valueBonferroni SignificancealphaModel=qwen7b, KV budge...Model=qwen7b, KV budget (b)=1282026.054.791.670.0022—alphaModel=llama8b, KV budg...Model=llama8b, KV budget (b)=642026.052.230.560.0054—alphaModel=qwen7b, KV budge...Model=qwen7b, KV budget (b)=642026.051.560.890.2—alphaModel=llama8b, KV budg...Model=llama8b, KV budget (b)=1282026.050.222.450.86—