safetyfactbullishGroup Relative Policy Optimization (GRPO) can incentivize models to reason by generating thinking traces prior to giving final answersComputer Vision02 Aug 2026http://arxiv.org/abs/2607.28464v1