CaptchaArena: A Large-Scale, Fine-Grained Dataset for Training Computer-Use Agents on Interactive CAPTCHAs
CaptchaArena gives agents 50,000 executable CAPTCHA-solving trajectories across 20 interactive puzzle types.
The paper says each solution is verified through execution, with 46,000 trajectories carrying step-by-step reasoning annotations. It also adds pixel-mask labels for irregular targets, aiming at finer supervision than prior CAPTCHA datasets. The authors train a 9B CaptchaAgent with supervised fine-tuning and reinforcement learning, reporting Pass@1 gains from 70.5 to 71.7 and improvements on two outside benchmarks. HF Daily Papers' note
The paper says each solution is verified through execution, with 46,000 trajectories carrying step-by-step reasoning annotations. It also adds pixel-mask labels for irregular targets, aiming at finer supervision than prior CAPTCHA datasets. The authors train a 9B CaptchaAgent with supervised fine-tuning and reinforcement learning, reporting Pass@1 gains from 70.5 to 71.7 and improvements on two outside benchmarks. HF Daily Papers' note
score 5