A new research framework called OpenVLHarness addresses a known limitation of multimodal large language models (MLLMs): their tendency to struggle with tasks requiring grounded perception, external knowledge, or precise numerical evidence that cannot be reliably produced through internal reasoning alone. The authors show that simply giving models more computation time ('inference budget') yields only marginal improvements across 22 benchmarks, suggesting that bigger reasoning chains are not a substitute for actual access to external evidence.
The proposed system is built around two main components. A capability layer standardizes access to heterogeneous external tools and converts their raw outputs into formats the model can interpret, effectively turning disparate backends into usable 'observations.' A persistent multimodal memory then stores images, structured data, and web references as addressable artifacts that remain available throughout a query, rather than being discarded after each reasoning step. A coding interface allows the model to write and execute programs over this stored data, enabling it to combine results from several tools in a single computation.
According to the paper, this architecture was tested across 22 benchmarks using eight different model backbones, both open-source and proprietary. The harness improved average scores by 9.7 to 12.6 points compared to equivalent models without tool access, with gains of 11.2 points for Qwen3-VL-8B, 9.7 points for Kimi K3, and 9.9 points for a model referred to as GPT-6 Astra. Notably, the authors report that a mid-effort configuration using another model variant, GPT-6 Luna, outperformed both the base model and competing approaches such as Codex and Visual Sketchpad run at maximum reasoning effort, while using at most 37% of their token cost.
The work fits into a broader trend in AI research: moving beyond pure chain-of-thought reasoning toward agentic systems that actively fetch, store, and manipulate external evidence. If the efficiency gains hold up under independent scrutiny, frameworks like this could make tool-augmented visual reasoning more practical for production use, since lower token consumption directly affects inference costs. As with many benchmark-driven papers, the real test will be whether these gains generalize beyond the specific backbones and tasks evaluated here.