VISTA: A Visual Harness for Reasoning in an Interactive World

TL;DR

VISTA is a visual harness that enables multimodal models to maintain long-horizon visual memory and actively reason over past observations in interactive environments. It significantly boosts performance on visual reasoning tasks, achieving perfect scores on ARC-AGI-3 with fewer actions than humans. The approach generalizes across diverse visual domains with minimal adaptation.

Why It Matters

This work demonstrates a scalable method to enhance multimodal agents' reasoning in dynamic visual environments, reducing reliance on task-specific engineering. It opens pathways for deploying general-purpose AI agents in real-world interactive settings with high efficiency.

Key Findings

• VISTA enables multimodal models to retain lossless visual memory of past observations, allowing re-examination and reorganization during reasoning.
• On ARC-AGI-3, VISTA achieves a perfect Relative Human Action Efficiency score of 100.00, outperforming Claude Opus 5.0 and using 57.4% fewer actions than humans.
• The framework generalizes across multiple visual benchmarks with minimal changes, consistently surpassing baselines using the same underlying model.