Tom McCoy Posts Paper on Emergent Symbolic Structure in Neural Networks

Tom McCoy, a Yale assistant professor of linguistics and computer science, posted a thread describing his arXiv paper. The work tests the hypothesis that neural networks succeed on symbolic tasks by building implicit symbolic structure in vector space using tensor product representations. It covers small models on list tasks and larger LLMs on math, logic, coding, and language. The approach approximates a network's representations with a TPR, then substitutes a closed-form TPR equation to check if outputs remain correct. Colleagues including Tal Linzen and Raphaël Millière commented on the thread, noting its relevance to interpretability and cognitive science.