ResearchBenchmarksCode Generation on BigCodeBench instruction split (test)Follow37.86Pass RateTEXTRESNET27.376830.098432.8235.5416Feb 9, 2026Evaluation ResultsMethodMethodLinksPass RateTEXTRESNETmode=deep residual tuningmode=deep residual tuning2026.0237.86TextGradsummarization=falsesummarization=false2026.0235.71TextGradsummarization=truesummarization=true2026.0235.12CoTmode=unoptimized lower...mode=unoptimized lower bound2026.0234.15DSPyoptimizer=MIPROoptimizer=MIPRO2026.0233.81HBCmode=hierarchical imit...mode=hierarchical imitation learning2026.0227.78