Deficient executive control in transformer attention

Area of Science:

• Cognitive Science
• Artificial Intelligence
• Neuroscience

Background:

• Transformers in LLMs utilize self-attention, revolutionizing NLP.
• Human attention has explicit executive control for conflict resolution and adaptive behavior.
• LLMs currently lack explicit executive control mechanisms found in human attention.

Purpose of the Study:

• To investigate the limitations of LLM attention mechanisms in conflict resolution.
• To compare LLM attention performance to human performance on the Stroop task.
• To assess the impact of context length on LLM attention and conflict resolution.

Main Methods:

• Utilized the classic color Stroop task to evaluate executive control of attention in LLMs.
• Tested LLMs on both congruent and incongruent conditions of the Stroop task.
• Varied the length of word lists to examine performance degradation under increasing interference.

Main Results:

• LLMs exhibited a conflict effect, with reduced accuracy in incongruent conditions for short word lists, mirroring human performance.
• Performance in incongruent conditions significantly degraded with increased word list length, approaching total collapse.
• Accuracy remained high in congruent conditions, and word reading accuracy was near-perfect across list lengths.

Conclusions:

• Transformer attention mechanisms in LLMs have fundamental limitations in conflict resolution over extended contexts.
• LLMs fail to adaptively up-regulate control under rising interference.
• Incorporating executive control mechanisms similar to biological attention is vital for advancing artificial general intelligence.