ResearchBenchmarksMathematical Reasoning on GSM8K JSON (test)Follow0.854Structural AccuracyRL-Struct0.123920.313460.5030.69254Nov 29, 2025Evaluation ResultsMethodMethodLinksStructural AccuracyJSON ValidityRL-StructAvg. Depth=2, learning...Avg. Depth=2, learning_paradigm=Reinforcement Learning2025.110.8540.889SFTAvg. Depth=2, learning...Avg. Depth=2, learning_paradigm=Supervised Fine-Tuning2025.110.5820.658Zero-shotAvg. Depth=2, learning...Avg. Depth=2, learning_paradigm=Zero-shot2025.110.1520.352