ResearchBenchmarksReasoning on AIME 24 (Accuracy, Improvement, Spikes, Repairs)Follow30Total ScoreBaseline28.529.253030.75Apr 20, 2026Evaluation ResultsMethodMethodLinksTotal ScoreCorrect CountAccuracyImprovement RateSpikes CountRepairs CountBaselineBase Model=Qwen3-8BBase Model=Qwen3-8B2026.04302376.7———SPREGBase Model=Qwen3-8BBase Model=Qwen3-8B2026.04302583.36.6819819