Kaiming He's Team Releases VISTA, a Visual-Native Harness Framework

Over the past year, Harness (scaffolding) has become an important path for improving agent capabilities. According to the source's description, Anthropic once used long-duration coding tasks as an example to explain how Harness helps a model push a task forward across multiple context windows: unfinished items are recorded in a task list, completed work is saved in progress files and code, and the model can continue by reading this information upon entering a new window. This mechanism mainly relies on text and code to save task state.

But the source points out that once an agent enters a real visual environment, text memory alone is not enough. The model must not only remember the position and orientation of objects, but also understand what changed in the environment before and after each action; moreover, when the model first sees an image, it may not know which detail will later become important. Such visual information is also hard to fully record in advance in a written note.

In benchmarks like ARC-AGI-3, an existing approach is to first convert frames into numeric text grids, then have the model write programs to simulate environment rules and verify action plans. The problem the source raises is this: the more complex the environment, the harder it is to fully convert object appearance, spatial relations, and dynamic changes into text and code; and the longer the interaction history, the more early frames get pushed out of context or replaced by text summaries. VISTA aims to answer exactly this gap—how, without additionally training a foundation model, to let an agent re-examine visual information it has seen before when reasoning requires it.

VISTA designs three core components: visual observation, lossless visual memory, and active visual inspection, responsible for letting the model see the frame clearly, saving history, and looking back on demand, respectively.

Visual observation is responsible for providing environment frames to the model. In the ARC-AGI-3 experiments, VISTA enlarged the official 64×64 frames into 512×512 PNG images while preserving the color, appearance, and spatial relations of objects.

Lossless visual memory is responsible for saving the frames the model sees. After each action is executed, VISTA fully saves all frames returned by the environment, including intermediate animation frames during the action, and indexes them by episode number and frame number. As the source puts it: the model does not need to judge in advance which information is worth remembering; it can first save the visual experience completely and leave it for later use.

Active visual inspection is responsible for looking back on demand. Through the inspect tool, the model can specify a historical episode to retrieve the corresponding frame, or crop and zoom into a local region to examine details; if it wants to know what a certain operation actually changed, it can also retrieve multiple frames at once to compare changes before and after the action. The source calls this mechanism explicit attention over interaction history—the model chooses, based on the question it is currently thinking about, which visual information re-enters the context.

Saving frames completely does not mean stuffing them all into context. In the full scheme, after each action the framework by default shows the model only the last frame, while intermediate frames during the action are uniformly stored in the visual archive and actively retrieved only when the model needs them. The source believes this both preserves complete visual information and avoids a large number of historical images continuously occupying context space.

VISTA also introduces two text notes: GUIDE.md records rules and experience reusable across different levels, and WORKING.md records the current level's state, progress, and follow-up plans. When context approaches its limit, the model first organizes a handoff summary, then enters a new context window to continue the task, with the previous text notes, action history, and visual archive all preserved.

According to the execution flow described by the source, at the start of each episode the model first observes the current frame and available actions, combines prior experience to judge the environment state and propose action hypotheses; when information is insufficient, it calls tools to review historical frames, zoom into local regions, or even read pixel information to find evidence; after finding evidence, it first predicts the changes the operation may bring, and after the action is executed compares the actual result with the prediction, thereby correcting its understanding of the rules.

The key constraint here is: VISTA does not additionally train a new model. Environment understanding, action planning, and reasoning are still done by existing multimodal models; the Harness is only responsible for executing tool calls, saving visual history, and providing evidence when needed. As the source summarizes, what VISTA changes is not the model itself, but the way the model acquires, stores, and uses visual information.

ARC-AGI-3 is an interactive visual reasoning benchmark that does not provide game rules or goals in advance; the agent must figure out how to clear the game through observation and action. The source says that with VISTA, Claude Opus 5.0 completed all 25 public ARC-AGI-3 games, reached a full score of 100 on relative human action efficiency, and used 57.4% fewer game actions than the first-play human baseline; switching to GPT-5.6 Sol, it also cleared all of them, with a score of 99.

The other three benchmarks cover browser games, mazes, and line-connection tasks. Using the same GPT-5.6 Sol model, compared with the official base framework, VISTA raised the success rate on GameWorld's 170 tasks from 40.0% to 63.3%; on AI GameStore's 10 games, the composite score rose from 47.3 to 140.3, with the human median normalized to 100; on the 39 maze and line-connection problems selected from BabyVision, accuracy rose from 41.0% to 63.2%. The source specifically notes that the last group of tasks contains only static images, and the model can still improve its judgments by zooming into local regions and inspecting pixels.

The source also says that the same VISTA framework requires only a small amount of adaptation to be used for different games and puzzles. In its conclusion, the research points future validation toward embodied tasks closer to the physical world: letting the model save, re-examine, and use its own visual experience in a continuously changing environment to decide the next action.

The level of evidence must first be made clear. The existence of the VISTA paper, the author composition, and the benchmark nature of ARC-AGI-3 are confirmed facts: the paper's three co-first authors are Qiushi Han, Hu Keya, and Linlu Qiu, and the other two authors are Cathy Wu and He Kaiming. All the evaluation results above, however, come from the paper's author team, relayed through a QbitAI report; they are source claims, with no third-party independent verification or reproduction.

One obvious explanatory gap is that the report does not state the baseline scores of Claude Opus 5.0 and GPT-5.6 Sol without VISTA. Therefore, the 100 and 99 on ARC-AGI-3 cannot be broken down to show how much VISTA contributed on its own. By contrast, the three data sets for GameWorld, AI GameStore, and BabyVision provide a comparison of "same model, official base framework versus VISTA," making the direction of the gain clearer, but the source likewise does not disclose the configuration details of these official base frameworks.

Several costs and boundaries also remain unanswered: the storage and retrieval overhead brought by the visual archive's long-term preservation of original frames is not explained by the source; whether the paper has been formally published or exists only as a preprint is also not made clear by the source; and VISTA's effectiveness on embodied tasks closer to the physical world has not yet been verified, with the team only listing it as a future research direction.

So the more prudent reading is this: the problem VISTA raises—how to make visual experience a context that can be called upon at any time and repeatedly examined—and the engineering path it offers (archive as-is, look back on demand, feed only the last frame by default) are the most noteworthy parts of this work; the specific scores should be regarded as team self-reported results awaiting reproduction, not as established capability conclusions.