Modeling Large Language Model Cognition

This project builds three core directions in a philosophy of cognitive science for large language models (LLMs): 1) interpreting LLM behavior, 2) analyzing the system in terms of Marr’s three levels, and 3) mapping the correspondence between behavioral outputs and the internals of the system. It consists of the following three chapters accordingly: Dennettian Stances and Large Language Models The aim of this chapter is to argue that there are convincing grounds for applying Dennett’s framework of stances to LLMs against worries that doing so produces “counterfeit people”. After introducing the intentional and design stances, I take sycophancy as a prototypical LLM behavior and consider how each stance might be used to describe and explain it. Having shown that neither stance alone is sufficient, I argue that the stances are best understood as explanatory and predictive strategies whose applicability to LLMs is pragmatic and epistemic, not metaphysical. That is, the question of whether to adopt the intentional or the design stance in a given context is a matter of explanatory choice, rather than an assessment of the intrinsic mental properties of LLMs. Three Levels for Large Language Model Cognition This chapter applies David Marr’s tri-partite analysis to large language models (LLMs). It offers a modified, pragmatist three-level framework for examining representative cases of explanations at the computational, algorithmic, and implementation levels. It seeks to make progress on the question of what kind of explanations would be necessary to effectively understand these systems. In response to that question, it argues that useful explanations are context-specific and that explanatory power does not necessarily increase when descending levels. The three-level approach should be perceived as a productive but leaky abstraction that nevertheless facilitates research for understanding and steering LLMs. The Complex Correspondence Between Behavior and Algorithm in Large Language Models Prominent research directions in large language model (LLM) interpretability assume a simple (1:1) correspondence between high-level behaviors, such as deception, and specific algorithmic structures within the neural network of the model. This, however, is not always warranted and perpetuates long-standing problems of operationalization in the cognitive sciences. The chapter instead proposes a complex correspondence view which models the behavior-algorithm relationship as a context-dependent and dynamic mapping. By drawing on the mechanistic modeling approach, it argues that complex cognitive phenomena in LLMs must be functionally decomposed into constituent sub-capabilities to be tractable for algorithmic analysis and highlights the need for non-human-centric cognitive ontologies to properly characterize these systems.