Back to News
Models
Language Models Share a Readout Bottleneck in Working Memory
Arxiv.org·October 15, 2026·1 min read
AI Summary
Large language models exhibit a human-like working memory limitation on the N-back task, showing decreased performance as memory load increases and making more errors on lure items that appear at nearby but incorrect positions. This suggests LLMs experience a similar "readout bottleneck" when accessing stored information, mirroring constraints observed in human cognitive processing.
From the source
Large language models (LLMs) reproduce a human working-memory signature on the N-back task: performance declines as memory load N increases, and errors concentrate on lures, items that repeat the current stimulus at a neighboring but incorrect serial position…
The full text couldn't be loaded here (the source may require a subscription).
View original at Arxiv.orgKeep reading
At Singapore launch, Anthropic defends its oversight of how customers use its model, citing AI’s ‘dual use’ risksFortune · 1d agoNVIDIA Expands Advanced Technology Access for Europe’s Scientific Community Through Open Learning Resources, Collaborations and ComputeNvidia.com · 1d agoWhat’s behind the BBC’s funding crisis? An expert explainsThe Conversation Africa · 1d agoGlean CEO Arvind Jain says he wants Glean to be easy to replaceThe Next Web · 1d ago
Was this useful?