Rule Extrapolation in Language Models: A Study of Compositional Generalization on OOD Prompts
This work addresses the challenge of compositional generalization in AI for researchers, but it is incremental as it builds on existing studies of OOD behavior in formal languages.
The paper tackled the problem of understanding how language models generalize to out-of-distribution (OOD) prompts by introducing rule extrapolation, a scenario where prompts violate compositional rules, and found that different architectures like Transformers and state space models vary in their ability to handle such extrapolation.
LLMs show remarkable emergent abilities, such as inferring concepts from presumably out-of-distribution prompts, known as in-context learning. Though this success is often attributed to the Transformer architecture, our systematic understanding is limited. In complex real-world data sets, even defining what is out-of-distribution is not obvious. To better understand the OOD behaviour of autoregressive LLMs, we focus on formal languages, which are defined by the intersection of rules. We define a new scenario of OOD compositional generalization, termed rule extrapolation. Rule extrapolation describes OOD scenarios, where the prompt violates at least one rule. We evaluate rule extrapolation in formal languages with varying complexity in linear and recurrent architectures, the Transformer, and state space models to understand the architectures' influence on rule extrapolation. We also lay the first stones of a normative theory of rule extrapolation, inspired by the Solomonoff prior in algorithmic information theory.