Multimodal models often struggle with long-horizon tasks because they encode visual observations only once, losing detail. VISTA is a simple visual harness that gives such models a lossless visual memory and the ability to actively retrieve and reorganize past observations. On the ARC-AGI-3 benchmark, VISTA improved Claude Opus 5.0's Relative Human Action Efficiency from 40.68 to a perfect 100.00, completing all 25 games using 57.4% fewer actions than first-time human participants. VISTA also outperformed baseline systems across three additional benchmarks covering diverse visual games and puzzles.
Visual memory harness helps multimodal agents solve long interactive tasks
VISTA lets a general-purpose model retrieve and reorganize original visual observations while reasoning across games and puzzles.
Academic
Qiushi Han · Keya Hu · Linlu Qiu · Cathy Wu · Kaiming He
Massachusetts Institute of Technology
Research Digest··2 min read
In one line
VISTA uses lossless visual memory to boost multimodal model performance to 100% on ARC-AGI-3.
Before you read this
Three ideas this paper builds onagent harnessA system that structures, verifies, monitors, or controls how an AI agent interacts with external tools, environments, or execution runs.186 papers →vision-language modelA model that handles tasks combining visual and textual information, such as editing images from conversation or reasoning across images and text.265 papers →agent memoryA repository where an agent stores past interactions and outcomes to inform future decisions.241 papers →
Read it as
Han et al. introduce VISTA, a model-agnostic harness designed to extend multimodal agents beyond their immediate visual context.
The authors equipped a general-purpose with tools for directly observing an interactive environment, retaining its complete visual history, and selecting previous observations for renewed inspection.
How it works
before and afterThe Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.