In plain words
Some picture questions need more than what a model can work out from the image alone. OpenVLHarness is a system that lets a visual agent call tools to find objects, search for information, or calculate measurements, then keep the results available for later steps.
Across 22 visual task datasets, Qwen3-VL-8B had a higher average score with the system than when answering directly without tools. That average does not mean it improved on every dataset.
The research
Some visual questions depend on evidence that a model may not reliably infer from its initial input, such as small objects, unfamiliar information, or measurements. Can an agent gather that evidence with tools and carry it forward so later steps can inspect and combine it?
OpenVLHarness gives visual agents consistent tool access and reusable multimodal memory so they can gather, inspect, and combine evidence across steps.
The authors tested OpenVLHarness against direct answers from the same Qwen3-VL-8B model across 22 datasets. Its average task score rose from 34.2 to 45.4, a gain of 11.2 points; the Qwen results average three runs.
Because the backbone and questions were matched, this comparison supports a benefit from using the harness and its tools across that dataset set. It is an average result, not evidence that every task improved.
The results cover the tested datasets and models. Qwen3-VL results average three runs, while proprietary-model results are single runs; the memory ablation also uses fixed-seed subsamples for larger datasets.
Failure diagnoses cover failed trajectories and do not show that proposed fixes would recover accuracy.