Symbolic and Language Agnostic Large Language Models
This addresses the problem of interpretability and inferential reliability in AI language models for researchers and practitioners, but it is incremental as it builds on existing bottom-up strategies.
The authors argue that the success of large language models stems from bottom-up reverse engineering at scale, but their subsymbolic nature hides knowledge in meaningless microfeatures and fails to capture inferential aspects of language; they propose creating symbolic, language-agnostic, and ontologically grounded large language models as an alternative.
We argue that the relative success of large language models (LLMs) is not a reflection on the symbolic vs. subsymbolic debate but a reflection on employing an appropriate strategy of bottom-up reverse engineering of language at scale. However, due to the subsymbolic nature of these models whatever knowledge these systems acquire about language will always be buried in millions of microfeatures (weights) none of which is meaningful on its own. Moreover, and due to their stochastic nature, these models will often fail in capturing various inferential aspects that are prevalent in natural language. What we suggest here is employing the successful bottom-up strategy in a symbolic setting, producing symbolic, language agnostic and ontologically grounded large language models.