The N-back task is a classic measure of working memory: a person must indicate when the current stimulus matches one presented N steps earlier. As N grows, human accuracy falls. According to a new preprint on arXiv, large language models show the same decline when given the same task, suggesting that they too struggle to retrieve information from a limited-capacity buffer.
The study also reports that LLMs make errors on "lures"—items that repeat the current stimulus at a neighboring position rather than the exact N-back position. This is a distinctive human error signature, often linked to interference during memory retrieval. The authors argue that both humans and LLMs face a "readout bottleneck": the bottleneck is not in storing information but in selecting the correct item when multiple similar traces compete.
The finding is notable because LLMs and biological brains differ greatly in architecture, yet they exhibit a comparable computational constraint. The abstract does not specify which models were tested or the exact task design, so the scope of the effect remains unclear. Still, the result offers a concrete way to compare working-memory behavior across artificial and natural systems, and it may inform how we evaluate model limitations on tasks that require maintaining context over time.