Report
A possible trade-off behind human and LLM working-memory limits
The presenter suggests shared representations support efficient learning and generalization but also create interference.
TLDR
In an October 6 post, the presenter proposed that humans and large language models may face the same computational trade-off behind working-memory limits: shared representations enable efficient learning and generalization but also create interference. They said they would present the work at COLM that day.
Combined views
—
1 Source, first seen ago
