ResearchBenchmarksReasoning on AIME 25 (Accuracy, Improvement, Spikes, Repairs)Follow30Total CountBaseline28.529.253030.75Apr 20, 2026Evaluation ResultsMethodMethodLinksTotal CountCorrect Answers CountAccuracyImprovement RateError Spike CountRepair CountBaselineBase Model=Qwen3-8BBase Model=Qwen3-8B2026.04301756.7———SPREGBase Model=Qwen3-8BBase Model=Qwen3-8B2026.04302376.720962962