Kaiming He's team unveils VISTA, a visual-native harness for multimodal agents
Kaiming He and collaborators at MIT describe VISTA, a visual-native harness for multimodal models, in the paper "VISTA: A Visual Harness for Reasoning in an Interactive World." Rather than retraining a base model, VISTA stores every raw frame returned by the environment outside the context and lets the model recall, crop, zoom into and re-inspect them when reasoning requires it. According to the report, on all 25 public ARC-AGI-3 games an agent paired with Claude Opus 5.0 reached a perfect score of 100 in action efficiency relative to humans, using 57.4% fewer actions than the first-play human baseline, while GPT-5.6 Sol also cleared every game with a score of 99. On GameWorld, AI GameStore and BabyVision the same GPT-5.6 Sol model improved success rate from 40.0% to 63.3%, raised AI GameStore score from 47.3 to 140.3 (human median normalized to 100) and lifted accuracy on 39 mazes and connect-the-dots tasks from 41.0% to 63.2%, with embodied tasks listed as future work.
If it holds up, a harness that changes only how visual information is stored and recalled—not the model itself—could be a low-cost route to stronger multimodal agents on long-horizon and embodied tasks.