CaptchaMind: RL-Based CAPTCHA Solver Achieves 82.9% Success Rate
A new benchmark called CaptchaBench has been launched by researchers, specifically aimed at large-scale training. It features 16,000 samples generated programmatically, spanning eight different task categories, complete with in-depth annotations at both the region and process levels. Evaluations indicated that current methods struggle with tasks that necessitate precise visual detail and regional comparisons. To tackle this issue, the team created CaptchaMind, a reinforcement learning-based solver that incorporates explicit reasoning process supervision, achieving an average success rate of 82.9% across the eight categories. This research is documented in arXiv paper 2605.19538.
Key facts
- CaptchaBench includes 16,000 programmatically generated samples across eight task categories.
- CaptchaBench has detailed region and process-level annotations.
- Existing methods fail on tasks requiring fine-grained visual detail capture and region-level comparison.
- CaptchaMind is an RL-based solver trained with explicit reasoning process supervision.
- CaptchaMind achieves 82.9% average success rate across eight categories.
- The paper is available on arXiv with ID 2605.19538.
Entities
—