A new MIT paper expands the team’s visual-memory system for AI beyond the interactive game benchmark in its earlier public demonstration, testing the same setup on browser games and static visual-tracking tasks. VISTA lets a model retrieve earlier screen images as it works instead of relying only on the latest frame or a compressed history.
VISTA stores every image returned by an environment in its original form, including intermediate animation frames. When the model needs to check what changed, it can retrieve a past frame or enlarge a region; another tool reads pixel colors. The model chooses when to inspect, then continues acting in the environment.
On 25 public ARC-AGI-3 interactive games, VISTA completed every game with Claude Opus 5.0 and GPT-5.6 Sol, scoring 100 and 99 on the paper’s Relative Human Action Efficiency (RHAE) measure. RHAE combines level completion and action efficiency: for a completed level, matching the first-time-human reference earns 100, using fewer actions can earn up to 115, and unfinished levels get zero. The overall score is not a percentage correct. The GPT result compares with 13.33 for the organizers’ minimal setup using the same model and maximum reasoning effort; Opus’s minimal-harness comparison used high reasoning effort, while VISTA used xhigh.
The Oct. 1 paper also tests GPT-5.6 Sol on 34 GameWorld browser games (170 tasks), 10 AI GameStore games and 39 BabyVision visual-tracking tasks. VISTA beat the same-model minimal-harness baseline on all three, the authors report; BabyVision accuracy rose from 41% to 63.2%. The team had already shared its ARC result on the VISTA project page in August, so the added benchmark evidence is the paper’s main new step.
The authors caution that all 25 ARC games are public, so exposure during model training cannot be ruled out. These evaluations cover games and static images, not physical robots or open-ended real-world interaction.