Large language models face challenges in understanding meaning rules
Rethinking Evaluation for Principled Generalization Using a Kripkean Challenge to Meaning as Use
Computation and Language
Summary
This paper discusses a problem with how large language models understand meaning based on how words are used. The authors argue that just looking at past examples can't tell a model the exact rules it should follow in new situations. They suggest current tests don’t fully show if models really understand language. Instead, the authors propose testing models on their ability to think abstractly and handle repeated rule applications to better measure real understanding.
What this means in practice
- •For natural language processing teams: Improve model testing by checking if language models keep concepts consistent when facing new types of texts.
- •For software engineers building ai assistants: Design evaluation methods that verify assistants truly understand user instructions beyond fixed examples.
A position paper. It proposes an approach and reports no results.
Authors
Niyar R Barman, Omar Naim, Nicholas Asher
Abstract
Large language models operationalize an intuitive, view of meaning as use. We argue that it faces a rule-following challenge. Using Kripke's interpretation of Wittgenstein, we argue that finite histories of use do not determine a unique rule for future extension. This reframes the debate about OOD generalization but also about NLP evaluation; benchmarks that measure performance on fixed use cases or local robustness, we argue, do not by themselves establish semantic competence. To evaluate whether models preserve rules and concepts under principled domain extension, we propose to evaluate two capacities, central to this ability: abstraction and recursion; and we offer an experimental paradigm for testing both.