According to @godofprompt, Tsinghua’s Cache-to-Cache lets models share KV-caches, improving accuracy up to 14.2% and speed 2.5x over text passing.

According to @godofprompt, Tsinghua shows extra overlapping or off-target skills slash agent success by up to 23 pts; a budgeted selector lifts wins.

According to @godofprompt, Tsinghua found agents rarely switch strategies, limiting gains despite higher execution scores across 1,338 runs.

According to @AINewsOfficial_ four AI milestones span humanoid agility, zero-shot tracking, dexterous robots, and an open 753B GLM 5.2 model.

According to The Rundown AI on X, researchers at Tsinghua University demonstrated a robot that rallies in tennis with human level consistency using real time perception and control. As reported by The Rundown AI, the system integrates high speed vision, trajectory prediction, and motion planning to position and swing a racket with timing precise enough for live rallies. According to Tsinghua University research communications cited by The Rundown AI, this performance suggests commercialization paths in autonomous sports training robots, embodied AI benchmarks for dynamic tasks, and industrial pick and place systems that require fast reaction under uncertainty.