VISTA: not a model, but a harness
ARC-AGI-3 is a set of 25 interactive puzzle games that test whether a model can probe and progress in unseen rules. The official baseline scores are unimpressive: the same Claude Opus 5.0 with no extra tooling scores 40.68, and GPT-5.6 with full reasoning scores 13.33. The VISTA framework, posted in early October by a five-person MIT team (including Kaiming He), rewrites these numbers as 100.00 and 99.00—a perfect run through all 25 games, with Opus 5.0 using a total of 7,302 actions versus a reference value of 17,135 for first-time human participants, i.e., 57.4% fewer.
VISTA is not a model; it is equipment wrapped around a model. The paper title says it plainly: *a visual harness*. It consists of three things.
1. Lossless visual memory
Every frame seen in the game is stored at original resolution, indexed by turn and frame number, and retrievable at any time. This is a rebellion against the mainstream approach—usually, old frames are deleted when the context fills up, or images are replaced by text summaries to save tokens. VISTA's stance: you don't know in advance which detail will become critical twenty turns later, so compression is a gamble. The paper has a real example of this: the model retrieves frames from the previous 22 turns at once to assemble a complete map.
2. Active visual inspection
When a small detail is hard to see, the model can initiate an inspect action: zoom in—there is a 13x magnification example in the paper, resolving which face of three small blocks carries purple markings to determine orientation—or read pixel values directly. It's the human gesture of squinting at the screen, turned into an interface.
3. Two note files
GUIDE.md records persistent rules of a game, reused across levels—for example, the rule confirmed at turn 79 that "clicking an orange block makes it disappear and leaves an orange X mark; clicking again rebuilds it." WORKING.md is a scratchpad for the current level. Reasoning itself uses free-form language, not programs. The paper's wording is cautious but firm: to their knowledge, this is the first vision-based system to achieve a perfect or near-perfect score on ARC-AGI-3 without program synthesis.
The most valuable discipline: reasoning is free, actions cost
The model can hypothesize, predict, and plan freely without consuming a single action; only executed moves count. The RHAE metric (Relative Human Action Efficiency) rewards exactly this—think it through, then act. The gap of 7,302 vs. 17,135 is not because the model got smarter, but because it is allowed to think before moving.
Context: two opposite bets on memory, in the same week
Placing this back into last week's context is telling. Recently, MemT-1.0 was examined here: it compresses a 128K history into a fixed 10M state, saving 66x storage—an extreme of the compression route. VISTA goes the opposite way to the end: not a single frame compressed, everything stored as-is. The difference lies in the task: for long conversational histories, redundancy is predictable; in game puzzles, where critical details appear is unpredictable, so dropping any frame is a gamble. Meanwhile, the two note files are memory in another form—raw frames are token-level retention, while GUIDE.md is distilled rule text. One system stands at both ends of the memory spectrum at once, with "inspection" as the bridge.
What the controls prove
The control numbers show how much the harness contributes: same models, both without programs—official baseline 40.68 vs. 13.33; with VISTA, 100.00 vs. 99.00. The sixty-plus-point gain comes entirely from the external machinery of memory, inspection, and notes. The models didn't change; the brain is the same brain.
No compression, no code, perfect scores—these are one philosophy: leave information fidelity to memory, rule organization to language, and reasoning to the model. Each step only lets the party best at it do the work.
Source notes
Paper: arXiv 2610.02200, "VISTA: A Visual Harness for Reasoning in an Interactive World" (2026-10-01; Qiushi Han and Keya Hu co-first authors; Linlu Qiu, Cathy Wu, Kaiming He; MIT). Code: github.com/joshhhhhan/VISTA. All figures (25 games, 40.68→100.00, 13.33, 99.00, 7,302 vs. 17,135 actions i.e. 57.4%, 22-turn map assembly, 13x zoom, the turn-79 orange-block rule in GUIDE.md, GameWorld 34 games + AI GameStore 10 games + BabyVision 39 tasks) come from the paper. The "first" claim follows the paper's wording "To our knowledge, the first vision-based system…without program synthesis." The perfect score is a tie ("matching other state-of-the-art systems," mostly contemporaneous work); "new record" phrasing has been corrected accordingly. The "57.4% fewer actions than human beginners" refers in the original to first-time human participants.