Deficient Executive Control in Transformer Attention

Abstract

Although transformers in the large language models (LLMs) effectively implement a self-attention mechanism that has revolutionized natural language processing, they lack an explicit implementation of executive control of attention found in humans which is essential for resolving conflicts and selecting relevant information in the presence of competing stimuli, and is critical for adaptive behavior. To investigate this limitation in LLMs, we employed the classic color Stroop task that is widely regarded as the gold standard for testing executive control of attention. Our results revealed a typical conflict effect of better performance in terms of accuracy in the congruent condition (e.g., naming the ink color of the word RED in red) compared to incongruent condition (e.g., naming the ink color of the word RED in blue), which is similar to human performance, in short sequences. However, as sequence length increased, the performance degraded toward chance levels on the incongruent trials despite maintaining excellent performance on congruent trials and near-perfect word reading ability. These findings demonstrate that while transformer attention mechanisms can achieve human-comparable performance in smaller contexts, they are fundamentally limited in their capacity for conflict resolution across extended contexts. This study suggests that incorporating executive control mechanisms akin to those in biological attention could be crucial for achieving more general reasoning and reliable performance toward artificial general intelligence.