Training AIs to Solve CAPTCHAs

Researchers have been exploring how to improve the ability of artificial intelligence models to solve visual CAPTCHAs. Currently, vision-language models can solve these challenges without specific training, but they do so from scratch each time, requiring the same amount of time and computational resources. Specialized detectors can solve CAPTCHAs quickly, but only for the categories they have been trained on. A new system has been developed that combines a fine-tuned detector with an open-weight vision-language model to improve performance. This system uses a confidence-based approach to route the solving process. The goal is to create a solver that improves with use, unlike existing models that do not learn from exposure to multiple instances of the same puzzle.