Alibaba research: ReVision cuts GUI agent costs; Agent-BRACE and anchored bipolicy self-play

Alibaba researchers published three notable agent-systems papers: ReVision exploits temporal visual redundancy across GUI screenshots to scale computer-use agents under fixed context budgets; Agent-BRACE decouples beliefs from actions in long-horizon partially-observable tasks; and an 'anchored bipolicy' self-play method exposes safety self-consistency gaps that standard red-teaming misses.