ResearchBenchmarksCode Correction on HumanEvalCode Correction on HumanEvalFollow25.7Pass@1CDLM15.31820.723.4Dec 17, 2025Evaluation resultsMethodMethodLinksPass@1CDLMRefinement steps=4, Co...Refinement steps=4, Confidence threshold=0.92025.1225.7BaseRefinement steps=4, Co...Refinement steps=4, Confidence threshold=0.92025.1216.8MDLMRefinement steps=4, Co...Refinement steps=4, Confidence threshold=0.92025.1215.7