VISTA: A Visual Harness for Reasoning in an Interactive World

The 30-second take

• What: VISTA is a visual harness that lets a multimodal model see an interactive environment, keep past frames losslessly, and retrieve or reorganize that visual memory while it reasons.
• Why it matters: Long-horizon visual agents today forget or compress what they saw; a reusable harness that unlocks existing models is a cheaper path than training a new eyes-and-memory stack from scratch.
• Who should care: Agent and multimodal researchers, evaluation designers around ARC-AGI-3, and product teams who wrap a frontier VLM rather than train one.

What the paper actually did

The authors argue that multimodal models already have strong reasoning and that the missing piece is a harness for interactive environments. VISTA gives a general-purpose multimodal model long-horizon vision: it perceives the environment through visual observations and keeps a lossless visual memory of past observations in their original form. The model can actively retrieve those observations and reorganize its visual input as it reasons. On ARC-AGI-3, VISTA raises Claude Opus 5.0’s Relative Human Action Efficiency from 40.68 to a perfect 100.00, and the model finishes all 25 public games using 57.4% fewer actions than first-time human participants. The same simple design is applied, with minimal adaptation, to three additional benchmarks of visual games and puzzles, where it substantially outperforms baselines that use the same underlying model with minimal harnesses.

What makes this disruptive

The scarce resource in interactive visual agents is often not another pretrained backbone — it is memory and control of what the model is allowed to look at. A lossless visual memory plus active retrieval is a small interface change with a large reported effect: 40.68 to 100 RHAE on ARC-AGI-3 and fewer actions than first-time humans across all 25 public games. Showing transfer to three other visual benchmarks with minimal adaptation is the generality claim. If the gain is really in the harness, then abundance comes from wrapping existing models rather than waiting for a new training run. That is a measurement-and-systems result, not a proof that every interactive job is solved.

Why it matters (outside the lab)

Abundance lens: expert visual problem-solving and long-horizon attention are still scarce human or expensive-agent services. If a simple harness unlocks already-trained multimodal models, capable interactive help can become a default software layer sooner than a new foundation-model cycle. Horizon is near-term for research agents and game-like environments, not a claim of a general workplace robot. Near-term: update how you scaffold VLMs. Medium-term: cost, reliability, and environments beyond puzzles decide whether this is a default.

Limitations & open questions

The standout number is on ARC-AGI-3 public games with Claude Opus 5.0; that is one model and one benchmark family. “Perfect 100” Relative Human Action Efficiency is a defined score, not a claim of human-level intelligence in the open world. Additional benchmarks are visual games and puzzles, which may not represent messy physical or workplace video. Lossless visual memory has an obvious context- and storage-cost question the abstract does not price. Baselines are same-model minimal harnesses — stronger alternative memories might close the gap. No product timeline. Preprint.