Stroop Test Reveals Core Limitation in Transformer Attention

Researchers tested frontier LLMs (GPT-5, Claude Opus 4.1, Gemini 2.5, GPT-4o) with the Stroop task and found their ability to inhibit automatic word-reading collapses as sequence length grows, with accuracy dropping sharply on longer or mixed lists. The results show transformer attention lacks sustained executive control compared to human cognition, revealing a fundamental architectural gap in long-context decision-making.