reasoningfactbullishREFACT improves grounding by training models to decide when reasoning steps need contextual grounding and at what granularity facts should be restatedComputation and Language28 Jul 2026http://arxiv.org/abs/2607.20833v1