MIT's VISTA clears all 25 public ARC-AGI-3 games, but the score won't count
What happened
An October 9 report says MIT's VISTA cleared all 25 public game environments in ARC-AGI-3, but public-test results don't count toward the official leaderboard, and the benchmark's authors reject what the result says about general intelligence. VISTA changed no model weights, mainly adding image representation, lossless history archiving and a review tool. A text-only system, AVO, also cleared every game with about 12% fewer actions than VISTA. The two differ in more than input modality, so the comparison is not a controlled ablation.
Written by AI from the coverage · updated 2 hours ago
Coverage
Follow the reports to see the story from different sides.
- Computing Life · Share · YageVISTA 在 ARC-AGI-3 公开关卡拿到满分，为何出题人不认可其通用智能含义
MIT 团队的 VISTA 在 ARC-AGI-3 的 25 个公开游戏环境中全部通关，但公开测试成绩不进入官方正式排行榜。系统未修改模型权重，主要增加图像表示、无损历史存档和回看工具；纯文本系统 AVO 同样全部通关，交互动作比 VISTA 少约 12%，但两者差异不止输入模态，不能视为受控消融实验。
Heat over time
Not enough continuous observations to draw a trend yet.