Benchmark,NextTerm-440M control,BestTerm original sweep,BestTerm saved Alice reproduction Ryskina greedy,30/57 (52.63%),38/57 (66.67%),38/57 (66.67%) Ryskina beam-4,32/57 (56.14%),40/57 (70.18%),"40/57 (70.18%), separate teacher sweep" OEIS-Eval-Neo,6555/19034 (34.4384%),6532/19034 (34.3175%),6536/19034 (34.3386%) M1 macro MAPE (lower better),17.623927,17.582548,17.813680 Polynomial arithmetic,94.2917%,94.5625%,94.5417% (4538/4800) Polynomial quadratic,86.3696%,86.3043%,86.2174% (3966/4600) Polynomial cubic,74.8409%,74.5682%,74.5682% (3281/4400) Polynomial quartic,68.1190%,67.9524%,68.1905% (2864/4200)