Language models can learn implicit multi-hop reasoning, but only if they have lots of training data

Abstract: Implicit reasoning is the ability of a LLM to solve multi-hop reasoning tasks in a single forward pass, without chain of thought. We investigate this capability using GPT2-style LLMs trained from scratch on controlled $k$-hop reasoning datasets ($k = 2, 3, 4$). We show that while such models can indeed learn implicit $k$-hop reasoning, the required training data grows exponentially in $k$, and the required number of transformer layers grows linearly in $k$. We offer a theoretical explanation for why this depth growth is necessary. We further find that the data requirement can be mitigated, but not eliminated, through curriculum learning.