Cognitive tasks originally developed for humans are now increasingly used to study language models. While applying these tasks is often straightforward, interpreting their results can be challenging. In particular, when a model underperforms, it is often unclear whether this results from a limitation in the cognitive ability being tested or a failure to understand the task itself. A recent study argues that GPT 3.5’s declining performance on 2-back and 3-back tasks reflects a working memory capacity limit similar to humans (Gong et al., 2024). By analyzing a range of open-source language models of varying performance levels on these tasks, we show that the poor performance is due at least in part to a limitation in task comprehension and task set maintenance. We challenge the best-performing model with progressively harder versions of the task (up to 10-back) and experiment with alternative prompting strategies, before analyzing model attentions. Our larger aim is to contribute to the ongoing conversation around refining methodologies for the cognitive evaluation of language models.

Analogical reasoning is a fundamental capacity of human cognition that allows us to reason abstractly about novel situations by relating them to past experiences. While it is thought to be essential for robust reasoning in AI systems, conventional approaches require significant training and/or hard-coding of domain knowledge to be applied to benchmark tasks. Inspired by cognitive science research that has found connections between human language and analogy-making, we explore the use of intuitive language-based abstractions to support analogy in AI systems. Specifically, we apply large pre-trained language models (PLMs) to visual Raven’s Progressive Matrices (RPM), a common relational reasoning test. By simply encoding the perceptual features of the problem into language form, we find that PLMs exhibit a striking capacity for zero-shot relational reasoning, exceeding human performance and nearing supervised vision-based methods. We explore different encodings that vary the level of abstraction over task features, finding that higher-level abstractions further strengthen PLMs’ analogical reasoning. Our detailed analysis reveals insights on the role of model complexity, in-context learning, and prior knowledge in solving RPM tasks.

We advance a novel explanation of similarity-based interference effects in subject-verb and reflexive pronoun agreement processing, grounded in surprisal values computed from a pretrained large-scale Transformer model, GPT-2. Specifically, we show that surprisal of the verb or reflexive pronoun predicts facilitatory interference effects in ungrammatical sentences, where a distractor noun that matches in number with the verb or pronouns leads to faster reading times, despite the distractor not participating in the agreement relation. We review the human empirical evidence for such effects, including recent meta-analyses and large-scale studies. We also show that attention patterns (indexed by entropy and other measures) in the Transformer show patterns of diffuse attention in the presence of similar distractors, consistent with cue-based retrieval models of parsing. But in contrast to these models, the attentional cues and memory representations are learned entirely from the simple self-supervised task of predicting the next word.

Processing difficulty in online language comprehension has been explained in terms of surprisal and entropy reduction. Although both hypotheses have been supported by experimental data, we do not fully understand their relative contributions on processing difficulty. To develop a better understanding, we propose a mechanistic model of perceptual decision making that interacts with a simulated task environment with temporal dynamics. The proposed model collects noisy bottom-up evidence over multiple timesteps, integrates it with its top-down expectation, and makes perceptual decisions, producing processing time data directly without relying on any linking hypothesis. Temporal dynamics in the task environment was determined by a simple finite-state grammar, which was designed to create the situations where the surprisal and entropy reduction hypotheses predict different patterns. After the model was trained to maximize rewards, the model developed an adaptive policy and both surprisal and entropy effects were observed especially in a measure reflecting earlier processing.