Visual harnesses may help AI models answer harder image questions

Harishkk.com reports that visual harnesses let vision-language models crop and inspect images while answering questions. On a 2,000-question benchmark, two Qwen models answered more questions correctly at least once across eight attempts with a harness, though their one-attempt scores dipped. GPT-6 Luna improved on both measures with the harness at medium reasoning effort and above, the authors report.