Peter M. Todd

Eric Lacosse, Mariana Duarte, Peter M. Todd et al.

Both humans and Large Language Models (LLMs) store a vast repository of semantic memories. In humans, efficient and strategic access to this memory store is a critical foundation for a variety of cognitive functions. Such access has long been a focus of psychology and the computational mechanisms behind it are now well characterized. Much of this understanding has been gleaned from a widely-used neuropsychological and cognitive science assessment called the Semantic Fluency Task (SFT), which requires the generation of as many semantically constrained concepts as possible. Our goal is to apply mechanistic interpretability techniques to bring greater rigor to the study of semantic memory foraging in LLMs. To this end, we present preliminary results examining SFT as a case study. A central focus is on convergent and divergent patterns of generative memory search, which in humans play complementary strategic roles in efficient memory foraging. We show that these same behavioral signatures, critical to human performance on the SFT, also emerge as identifiable patterns in LLMs across distinct layers. Potentially, this analysis provides new insights into how LLMs may be adapted into closer cognitive alignment with humans, or alternatively, guided toward productive cognitive \emph{disalignment} to enhance complementary strengths in human-AI interaction.

Jared Lorince, Sam Zorowitz, Jaimie Murdock et al.

A folksonomy is ostensibly an information structure built up by the "wisdom of the crowd", but is the "crowd" really doing the work? Tagging is in fact a sharply skewed process in which a small minority of "supertagger" users generate an overwhelming majority of the annotations. Using data from three large-scale social tagging platforms, we explore (a) how to best quantify the imbalance in tagging behavior and formally define a supertagger, (b) how supertaggers differ from other users in their tagging patterns, and (c) if effects of motivation and expertise inform our understanding of what makes a supertagger. Our results indicate that such prolific users not only tag more than their counterparts, but in quantifiably different ways. Specifically, we find that supertaggers are more likely to label content in the long tail of less popular items, that they show differences in patterns of content tagged and terms utilized, and are measurably different with respect to tagging expertise and motivation. These findings suggest we should question the extent to which folksonomies achieve crowdsourced classification via the "wisdom of the crowd", especially for broad folksonomies like Last.fm as opposed to narrow folksonomies like Flickr.