ResearchBenchmarksCode Generation on Hard Corpus 4 (192 problems)Follow75.5S-optAlgoSkill40.34849.47458.667.726Jun 29, 2026Evaluation ResultsMethodMethodLinksS-opt∆ASAlgoSkillBackbone=GPT-5.5, Judg...Backbone=GPT-5.5, Judge Model=Haiku judge2026.0675.55.7DirectBackbone=GPT-5.5, Judg...Backbone=GPT-5.5, Judge Model=Haiku judge2026.0669.8—AlgoSkillBackbone=gpt-oss-120b,...Backbone=gpt-oss-120b, Judge Model=Haiku judge2026.0657.315.6DirectBackbone=gpt-oss-120b,...Backbone=gpt-oss-120b, Judge Model=Haiku judge2026.0641.7—