TuringSpace Benchmark: Top MLLMs Still Fail 26% of Interactive CAPTCHAs

Researchers from Penn, UMN, USC, Meta and MBZUAI introduced TuringSpace, a benchmark of 125 interactive CAPTCHAs testing five spatial abilities, where humans score 94.1% but top MLLMs like GPT-6 still fail about 26% of the tasks.

2026-11-20 ~ 2026-11-21 · 2 related posts