OpenVLHarness, a preprint submitted on October 19, 2026, puts a tool layer and multimodal memory around existing models. Its authors report average gains of 9.7–12.6 points across 22 benchmarks against matched versions without tools. These are the research team’s findings; SIExpo has not independently reproduced them.

What the memory retains

The public repository’s documentation says images, structured data and web references remain available within a question; the next question starts with fresh memory. Available tools include OCR, zoom, search and Python. Generated code runs with the user’s permissions and filesystem and network access; its timeout is not a sandbox. The main code uses MIT, while external models, services and data have separate terms.

Our explanation of the approach is straightforward: reading an image, finding an additional fact and calculating a relationship are different operations. Retaining an intermediate result lets the next step work on it. That does not remove the possibility that the initial observation is wrong. A calculation using an incorrectly read number can be mathematically consistent and still produce the wrong answer.

What a benchmark improvement establishes

The score difference provides a reason to examine how the tools are organized. It does not establish equal benefits for every document, language or business task. Our assessment is that a useful comparison should preserve the underlying model and expose what changes when tools are added. Changing the model, input data and budget simultaneously would make the contribution of the orchestration difficult to separate.

For an illustrative invoice evaluation, we would examine line recognition, selection of the relevant amount and arithmetic separately. This is an editorial example, not a published project test. Retained intermediate evidence could help locate the step where a discrepancy arises instead of leaving reviewers to judge only the final prose. The expected answer and original document should remain available to the person reviewing that discrepancy.

A public repository makes the approach available for investigation, but preparing the execution environment remains part of that work. Before running it, we would separate test files, restrict permissions and record the services involved. This would let an evaluation describe answer quality alongside the cost and conditions of obtaining it. Successful tool use and a safe deployment environment need their own evidence rather than being inferred from a benchmark gain.

Editorial verification: 2026-10-20T16:10:48.620127+00:00 (UTC). Cover: conceptual AI illustration, not a documentary photograph.