Researchers have introduced TuringSpace, a novel environment featuring 125 interactive CAPTCHAs designed to test and train multimodal large language models (MLLMs) in spatial intelligence. Unlike previous benchmarks, TuringSpace requires models to perform actions like clicking and dragging within a synthetic web environment, mimicking human interaction with visual and spatial information. While humans solve 94.1% of these tasks, the leading model, GPT-6 Astra, only achieves 74.1%, highlighting MLLMs' struggles with spatial reasoning and correcting their own actions. Fine-tuning the Qwen3.5-9B model on TuringSpace significantly improved its performance, demonstrating the environment's potential for teaching and evaluating spatial abilities that transfer to real-world scenarios. AI
IMPACT This benchmark highlights current limitations in MLLMs' spatial reasoning and provides a new method for training and evaluating these capabilities.
RANK_REASON The cluster describes a new benchmark and research paper for evaluating MLLMs.
Read on Hugging Face Daily Papers →
AI-generated summary · Google Gemini · from 2 sources. How we write summaries →