@kazani
VISTA: a visual harness that gives a VLM long-horizon vision for reasoning in an interactive world. With Claude Opus 5.0, it reaches 100% RHAE on ARC-AGI-3, perfectly solving all 25 public games. Its central intervention is essentially: Let the model see the world naturally, preserve everything it has seen, and let it deliberately look back. https://vista-research.github.io
0 reply
0 recast
0 reaction