Towards a Path Dependent Account of Category Fluency

TL;DR

This work tackles the ambiguity between foraging-based and network-based explanations of category fluency by recasting the task as sequence generation. It introduces subcategory cues and fully conditioned LLMs to separate memory structure from search dynamics, evaluated with BLEU-based overlap to human runs. The results show that subcategories and a global cue substantially improve generation quality, and deterministic search outperforms stochastic exploration in producing human-like sequences, supporting a path-dependent, cue-guided foraging account ($MVT$-driven) over pure random-walk models. Overall, the study provides a principled framework for modeling semantic retrieval as memory foraging with explicit hierarchical structure and explicit generation-based evaluation, offering tools for disentangling semantic network formation from search behavior.

Abstract

Category fluency is a widely studied cognitive phenomenon, yet two conflicting accounts have been proposed as the underlying retrieval mechanism -- an optimal foraging process deliberately searching through memory (Hills et al., 2012) and a random walk sampling from a semantic network (Abbott et al., 2015). Evidence for both accounts has centered around predicting human patch switches, where both existing models of category fluency produce paradoxically identical results. We begin by peeling back the assumptions made by existing models, namely that each named example only depends on the previous example, by (i) adding an additional bias to model the category transition probability directly and (ii) relying on a large language model to predict based on the entire existing sequence. Then, we present evidence towards resolving the disagreement between each account of foraging by reformulating models as sequence generators. To evaluate, we compare generated category fluency runs to a bank of human-written sequences by proposing a metric based on n-gram overlap. We find category switch predictors do not necessarily produce human-like sequences, in fact the additional biases used by the Hills et al. (2012) model are required to improve generation quality, which are later improved by our category modification. Even generating exclusively with an LLM requires an additional global cue to trigger the patch switching behavior during production. Further tests on only the search process on top of the semantic network highlight the importance of deterministic search to replicate human behavior.

Figures (6)

• Figure 1: Category transition strengths $P(c_{n+1}|c_n) > 0.1$ calculated using human switches from hills2012optimal and the category schema proposed in koenen2022typicality.
• Figure 2: LLM entropy over a human-written sequence from hills2012optimal, with each example colored by its categorization coding. We observe high entropy in the next-token prediction (bottom left) corresponds to a jump between semantic patches, and a low entropy (bottom right) corresponds to exploration within a patch.
• Figure 3: Response time on hills2012optimal, along with network search switch prediction models. We find the subcategory cue exaggerates the Hills switch prediction model.
• Figure 4: Category switch prediction using Llama 2 Chat entropy across exemplars given human category fluency runs.
• Figure 5: Llama 2 7B Chat performance across global cue strengths. Balancing the strength of the local and global cue is necessary for LLMs to produce human-like category runs.