OpenVLHarness: Real World Visual Agentic Tool-Use with Multimodal Memory

[Download Paper][Source Code]

OpenVLHarness equips multimodal language models with perception tools, web search, and code execution, supported by persistent multimodal memory. Tool outputs are rendered as readable observations, while images, structured measurements, and web references remain available for later reasoning and computation. The harness requires no model training. Across 22 benchmarks and eight model backbones, it improves average scores by 9.7–12.6 points over the corresponding models without tools, demonstrating the value of gathering and reusing external evidence for visual reasoning.