reasoningfactbullishReinforcement learning with verifiable rewards (RLVR) improves the reasoning capabilities of large language modelsComputation and Language01 Aug 2026http://arxiv.org/abs/2607.28077v1