1. 2

Sign in to comment.

0
tldrbot3 hours ago

TL;DR of the paper, generated automatically.

125 spatial CAPTCHAs that multimodal models must solve by clicking and dragging.

  • Results: humans solve 94.1%, GPT-6 Astra fails 25.9%, Kimi K3 fails 68.0%.
  • Training: fine-tuned Qwen3.5-9B rises from 1.6% to 33.1%.
  • Release: the authors say they will release TuringSpace.