OpenVLHarness
A multimodal reasoning harness that converts external evidence into addressable artifacts through a unified interface for heterogeneous tools, evidence rendering, and persistent multimodal memory, enabling visual agents to acquire, inspect, and combine evidence during multi-step reasoning.
Timeline (last 90 days)
The OpenVLHarness team conducted controlled experiments on 22 benchmarks and found that simply increasing the inference budget of the base model yields limited gains.
OpenVLHarness is a multimodal reasoning harness whose core consists of two parts: a capability layer and persistent multimodal memory.
The capability layer combines task-level tool interfaces with evidence rendering, converting tool return results into an observation format readable by the model.
Persistent multimodal memory retains images, structured data, and web citations throughout the entire query process, storing them as addressable artifacts.
OpenVLHarness provides a coding interface that allows the model to execute generated programs against stored data.
In evaluations across 22 benchmarks and eight open-source and proprietary backbone models, OpenVLHarness improves by an average of 9.7–12.6 points over matched tool-free baselines.
Qwen3-VL-8B improves by 11.2 points using OpenVLHarness.
Kimi K3 improves by 9.7 points using OpenVLHarness.
Running OpenVLHarness with GPT-6 Luna at medium reasoning effort, its performance exceeds that of the base model, Codex, and Visual Sketchpad at the highest reasoning effort, leading by 7.7–8.5 points.
All Facts (9)
The OpenVLHarness team conducted controlled experiments on 22 benchmarks and found that simply increasing the inference budget of the base model yields limited gains.
50%UnverifiedOpenVLHarness is a multimodal reasoning harness whose core consists of two parts: a capability layer and persistent multimodal memory.
50%UnverifiedThe capability layer combines task-level tool interfaces with evidence rendering, converting tool return results into an observation format readable by the model.
50%UnverifiedPersistent multimodal memory retains images, structured data, and web citations throughout the entire query process, storing them as addressable artifacts.
50%UnverifiedOpenVLHarness provides a coding interface that allows the model to execute generated programs against stored data.
50%UnverifiedIn evaluations across 22 benchmarks and eight open-source and proprietary backbone models, OpenVLHarness improves by an average of 9.7–12.6 points over matched tool-free baselines.
50%UnverifiedQwen3-VL-8B improves by 11.2 points using OpenVLHarness.
50%UnverifiedKimi K3 improves by 9.7 points using OpenVLHarness.
50%UnverifiedRunning OpenVLHarness with GPT-6 Luna at medium reasoning effort, its performance exceeds that of the base model, Codex, and Visual Sketchpad at the highest reasoning effort, leading by 7.7–8.5 points.
50%