SolveBind preprint reports gains in agents' CAPTCHA completion

Researchers report in an arXiv preprint that SolveBind, a framework for agents operating graphical interfaces, outperformed the strongest baseline by 13.3-34.5 percentage points across eight of nine tested backbone models on CAPTCHA tasks. The framework breaks verification into linked steps and supervises execution using evidence from tools, adapting actions when runtime feedback changes. The tests distinguish solving a challenge from completing interactions that anti-automation systems accept. In a paired GPT-5.5 analysis, SolveBind completed seven of 13 tasks where MobileAgent answered correctly but failed dynamic verification. Removing recovery limits raised success from 82.22% to 83.56%, but increased latency and token usage to 3.7x and 2.4x their full-supervision values.