Digital Agent outperforms o1 by 15% – trained with new RL-variant similar to R1arxiv.org·11 pts·let_tim_cook_·0