Native Parallel Reasoner: Reasoning in Parallelism via Self-Distilled Reinforcement Learning

Benchmark Model Rank Results
logic-puzzle-solving-on-zebralogicNPR (Qwen3-4B-Instruct-2507)#6Accuracy: 81.7
mathematical-reasoning-on-minerva-mathNPR (Qwen3-4B)#14Accuracy: 47.1
mathematical-reasoning-on-olympiadbenchNPR (Qwen3-4B-Instruct-2507)#15Accuracy: 63.7
mathematical-reasoning-on-math-500NPR (Qwen3-4B-Instruct-2507)#41Accuracy: 93.6
mathematical-reasoning-on-aime24NPR (Qwen3-4B-Instruct-2507)#62Acc: 63.3
mathematical-reasoning-on-aime-2025NPR (Qwen3-4B)#79Accuracy: 53.8