Tom McCoy posted a thread describing his paper that tests whether neural networks succeed on symbolic tasks by building implicit structure via tensor product representations. He analyzes small models on list tasks and larger models on math, logic, coding, and language. The approach approximates network representations with a TPR and checks whether feeding that approximation back into the network still yields correct outputs. McCoy links the work to a 1990 hypothesis by Paul Smolensky. Replies from researchers including Tal Linzen and Peng Qi note connections to earlier ideas in the field.