Holistic Representation in Large Language Models

This paper investigates the extent to which vector space representations in large language models (LLMs) can be understood as holistic representations. It is argued that holistic principles influence both the training process and the generation of output in LLMs. To contextualize this claim, debates on representation in psychology and philosophy will be presented, with a focus on the distinction between indicator and structural representations which recently emerged in philosophy of cognitive science. The holistic properties of structural representations are introduced as a key feature that distinguishes them from indicator representations and provides a basis for analyzing holism in LLMs. Two novel criteria for holism are developed: holistic perturbability and holistic exploitability. It is argued that LLMs meet both criteria. Moreover, holistic exploitation in output generation provides a novel way to understand the exploitability of structural representations in general, possibly also applicable in the human brain. Classic counterarguments to holism and their implications for the question of semantic grounding are discussed. Finally, implications for the homunculus problem, gestalt perception, and development are presented.