MIT agent VISTA uses 57.4 percent fewer steps than humans on puzzles
VISTA keeps all image frames in memory, and its score with GPT-5.6 Sol was 70, and not 99, without this.
Qiushi Han and colleagues at MIT released VISTA, control software that runs an AI model as an agent. On ARC-AGI-3, a test of 25 visual puzzle games, VISTA used 57.4 percent fewer steps than humans who saw the puzzles for the first time. It also did better than humans on 2 more game tests, GameWorld and AI GameStore. It is free with the MIT license.
How it works
VISTA works with Anthropic Claude Opus 5.0 or OpenAI GPT-5.6 Sol. It keeps all image frames, steps and results. The model can read 2 sets of notes, get kept frames, zoom in on a frame, or make a step.
The numbers
- ARC-AGI-3: Claude Opus 5.0 scored 100 and GPT-5.6 Sol scored 99.
- GameWorld: GPT-5.6 Sol completed 63.3 percent of tasks. Humans completed 55.3 percent.
- Without frame storage, GPT-5.6 Sol scored 70.
What is not known
The authors tested only games. The source gives no data on cost or speed, and no test on business tasks. Tycho also scored 100, but it made code that copied the games.
Insight: The memory method of an agent decides results
Vercel Labs reads screen item records and not pixels. VISTA keeps all frames. A different study found that tool graph memory gave small increases and used more model calls. The memory method, not the model, seems to be where agent builders compete. Plain stores with a find tool can gain. Test the 2 methods on your task.
Sources
Posted
