Tom McCoy on Compositional Filler-Role Representations

Tom McCoy, assistant professor of linguistics at Yale, posted thread replies addressing how gradient and statistical information is encoded and used by LLMs. He notes it is almost certainly encoded in many domains but not in the fully systematic tasks studied. One reply states hypothesis 5 that the filler-role representations are compositional, enabling generalization to novel filler-role combinations. McCoy adds that this could have been false and references appendix M, which contrasts two networks that both perform nearly perfectly on their task yet differ in showing systematic filler-role composition.