ResearchBenchmarksMathematical Reasoning on DeepMath 103K subsetFollow65.9AccuracyTTVS43.22849.1145560.886Apr 9, 2026Evaluation ResultsMethodMethodLinksAccuracyTTVSBackbone=Qwen3-4BBackbone=Qwen3-4B2026.0465.9TTRLBackbone=Qwen3-4BBackbone=Qwen3-4B2026.0462.8InitBackbone=Qwen3-4BBackbone=Qwen3-4B2026.0453.2TTVSBackbone=LLaMA-3.2-3B-...Backbone=LLaMA-3.2-3B-Instruct2026.0449TTRLBackbone=LLaMA-3.2-3B-...Backbone=LLaMA-3.2-3B-Instruct2026.0446.6InitBackbone=LLaMA-3.2-3B-...Backbone=LLaMA-3.2-3B-Instruct2026.0444.1