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

A controlled study across 22 benchmarks found that simply increasing the base multimodal large language model's inference budget yields only marginal gains on visual reasoning tasks requiring external evidence, according to research introducing OpenVLHarness, a real-world visual agentic tool-use system with multimodal memory. OpenVLHarness addresses this by giving visual agents tools and multimodal memory rather than relying on internal reasoning alone. The work matters for developers building visual agents that must gather evidence beyond what a model can reason out internally.

Multimodal large language models MLLMs have made rapid progress on visual reasoning, yet many tasks require evidence that is difficult to obtain through internal reasoning alone: our controlled study on 22 benchmarks shows that simply increasing the inference budget of the base model yields only m