Transformer attention has a working memory problem
Researchers found that transformer attention mechanisms lack the executive control functions that let human brains manage working memory. The models can retrieve information, but they cannot suppress irrelevant context.
A team from Oxford published a comparative study showing that transformer attention mechanisms are missing a critical piece of how human working memory actually works. They mapped attention heads to three cognitive functions: maintenance, retrieval, and executive control. The first two are there. The third is not. Executive control is the ability to suppress irrelevant information when you need to focus. Human brains do this constantly. If you are recalling a phone number while someone is talking next to you, your prefrontal cortex actively dampens the noise. Transformers do not have an equivalent mechanism. They retrieve everything that looks relevant by similarity, but they cannot turn down the volume on what should be ignored. This explains a pattern I have seen in LLM outputs where the model fixates on a keyword from earlier in the conversation even when the topic has shifted. It is not that the model forgot the earlier context. It is that it has no mechanism to decide that the earlier context should now be backgrounded. Attention weights can only be set based on what is relevant, not what should be suppressed. The researchers tested this by giving transformers tasks where success required ignoring distractors. The models failed more often than humans on tasks where executive control mattered. On pure retrieval tasks, they matched or beat human performance. This is a neat framing for why RAG systems break when the retrieved documents contain misleading but topically similar content. The model has no executive gate to decide that Document 3 is a red herring even if it shares vocabulary with the query. It weights everything by similarity and hopes for the best. I wonder if this is fixable with architecture changes or if it requires a fundamentally different attention mechanism. Suppression is not the same operation as weighting by relevance. It might need an adversarial signal or a second attention pass that explicitly models what to ignore.
Source: Deficient executive control in transformer attention