reasoningfactneutralSFT made the model shorter and more consistently Turkish in reasoning behavior but results were mixed, with large reductions in response lengthComputation and Language27 Jul 2026http://arxiv.org/abs/2607.01927v1