agentsfactneutralDirect supervised fine-tuning on collected experience recovers significantly less performance than Experience DistillationComputation and Language28 Jul 2026http://arxiv.org/abs/2607.21051v1