OpenVLHarness adds tool use and memory for visual agent reasoning
Original title: OpenVLHarness: Real World Visual Agentic Tool-Use with Multimodal Memory
On 22 benchmarks it lifts matched no-tool baselines by 9.7 to 12.6 points across eight backbones.
OpenVLHarness is a multimodal reasoning harness with a capability layer and persistent multimodal memory for visual agentic tool use. It renders heterogeneous tool outputs as model-readable observations and stores images, structured data, and web references as addressable artifacts. A coding interface runs generated programs over stored data. On 22 benchmarks with eight open and proprietary backbones, average scores rose 9.7-12.6 points over matched no-tool baselines.
- Raising the base model's inference budget alone gave only modest gains.
- Qwen3-VL-8B gained 11.2 points, Kimi K3 9.7, and GPT-6 Astra 9.9.
- With GPT-6 Luna at medium effort it beat max-effort baselines by 7.7-8.5 points.
- That run used at most 37 percent of those baselines' model-token cost.
From the original article
1 min read at huggingface.co
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 modest gains.
By Yu Zhou, Cheng-Fu Yang, Rui Sun · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.