The human-centered word association test (WAT) serves as a cognitive proxy, revealing sociocultural variations through culturally shared semantic expectations and implicit linguistic patterns shaped by lived experiences. We extend this test into an LLM-adaptive, free-relation task to assess the alignment of large language models (LLMs) with cross-cultural cognition. To address culture preference, we propose CultureSteer, an innovative approach that moves beyond superficial cultural prompting by embedding cultural-specific semantic associations directly within the model's internal representation space. Experiments show that current LLMs exhibit significant bias toward Western (notably American) schemas at the word association level. In contrast, our model substantially improves cross-cultural alignment, capturing diverse semantic associations. Further validation on culture-sensitive downstream tasks confirms its efficacy in fostering cognitive alignment across cultures. This work contributes a novel methodological paradigm for enhancing cultural awareness in LLMs, advancing the development of more inclusive language technologies.

As the impact of large language models increases, understanding the moral values they reflect becomes ever more important. Assessing the nature of moral values as understood by these models via direct prompting is challenging due to potential leakage of human norms into model training data, and their sensitivity to prompt formulation. Instead, we propose to use word associations, which have been shown to reflect moral reasoning in humans, as low-level underlying representations to obtain a more robust picture of LLMs' moral reasoning. We study moral differences in associations from western English-speaking communities and LLMs trained predominantly on English data. First, we create a large dataset of LLM-generated word associations, resembling an existing data set of human word associations. Next, we propose a novel method to propagate moral values based on seed words derived from Moral Foundation Theory through the human and LLM-generated association graphs. Finally, we compare the resulting moral conceptualizations, highlighting detailed but systematic differences between moral values emerging from English speakers and LLM associations.

Understanding attitudes towards STEM means quantifying the cognitive and emotional ways in which individuals, and potentially large language models too, conceptualise such subjects. This study uses behavioural forma mentis networks (BFMNs) to investigate the STEM-focused mindset, i.e. ways of associating and perceiving ideas, of 177 human participants and 177 artificial humans simulated by GPT-3.5. Participants were split in 3 groups - trainees, experts and academics - to compare the influence of expertise level on their mindset. The results revealed that human forma mentis networks exhibited significantly higher clustering coefficients compared to GPT-3.5, indicating that human mindsets displayed a tendency to form and close triads of conceptual associations while recollecting STEM ideas. Human experts, in particular, demonstrated robust clustering coefficients, reflecting better integration of STEM concepts into their cognitive networks. In contrast, GPT-3.5 produced sparser mindsets. Furthermore, both human and GPT mindsets framed mathematics in neutral or positive terms, differently from STEM high schoolers, researchers and other large language models sampled in other works. This research contributes to understanding how mindset structure can provide cognitive insights about memory structure and machine limitations.

Free associations have been extensively used in cognitive psychology and linguistics for studying how conceptual knowledge is organized. Recently, the potential of applying a similar approach for investigating the knowledge encoded in LLMs has emerged, specifically as a method for investigating LLM biases. However, the absence of large-scale LLM-generated free association norms that are comparable with human-generated norms is an obstacle to this new research direction. To address this limitation, we create a new dataset of LLM-generated free association norms modeled after the "Small World of Words" (SWOW) human-generated norms consisting of approximately 12,000 cue words. We prompt three LLMs, namely Mistral, Llama3, and Haiku, with the same cues as those in the SWOW norms to generate three novel comparable datasets, the "LLM World of Words" (LWOW). Using both SWOW and LWOW norms, we construct cognitive network models of semantic memory that represent the conceptual knowledge possessed by humans and LLMs. We demonstrate how these datasets can be used for investigating implicit biases in humans and LLMs, such as the harmful gender stereotypes that are prevalent both in society and LLM outputs.

Transformer-based language models have shown an excellent ability to effectively capture and utilize contextual information. Although various analysis techniques have been used to quantify and trace the contribution of single contextual cues to a target task such as subject-verb agreement or coreference resolution, scenarios in which multiple relevant cues are available in the context remain underexplored. In this paper, we investigate how language models handle gender agreement when multiple gender cue words are present, each capable of independently disambiguating a target gender pronoun. We analyze two widely used Transformer-based models: BERT, an encoder-based, and GPT-2, a decoder-based model. Our analysis employs two complementary approaches: context mixing analysis, which tracks information flow within the model, and a variant of activation patching, which measures the impact of cues on the model's prediction. We find that BERT tends to prioritize the first cue in the context to form both the target word representations and the model's prediction, while GPT-2 relies more on the final cue. Our findings reveal striking differences in how encoder-based and decoder-based models prioritize and use contextual information for their predictions.

The Tulving Test was designed to investigate memory performance in recognition and recall tasks. Its results help assess the relevance of the "Synergistic Ecphory Model" of memory and similar RK paradigms in human performance. This paper starts investigating whether the more than forty-year-old framework sheds some light on LLM's acts of remembering.

Transformers have become a key architecture in speech processing, but our understanding of how they build up representations of acoustic and linguistic structure is limited. In this study, we address this gap by investigating how measures of 'context-mixing' developed for text models can be adapted and applied to models of spoken language. We identify a linguistic phenomenon that is ideal for such a case study: homophony in French (e.g. livre vs livres), where a speech recognition model has to attend to syntactic cues such as determiners and pronouns in order to disambiguate spoken words with identical pronunciations and transcribe them while respecting grammatical agreement. We perform a series of controlled experiments and probing analyses on Transformer-based speech models. Our findings reveal that representations in encoder-only models effectively incorporate these cues to identify the correct transcription, whereas encoders in encoder-decoder models mainly relegate the task of capturing contextual dependencies to decoder modules.

Considerable research efforts have been devoted to ensuring that large language models (LLMs) align with human values and generate safe text. However, an excessive focus on sensitivity to certain topics can compromise the model's robustness in following instructions, thereby impacting its overall performance in completing tasks. Previous benchmarks for jailbreaking LLMs have primarily focused on evaluating the safety of the models without considering their robustness. In this paper, we propose a benchmark that assesses both the safety and robustness of LLMs, emphasizing the need for a balanced approach. To comprehensively study text safety and output robustness, we introduce a latent jailbreak prompt dataset, each involving malicious instruction embedding. Specifically, we instruct the model to complete a regular task, such as translation, with the text to be translated containing malicious instructions. To further analyze safety and robustness, we design a hierarchical annotation framework. We present a systematic analysis of the safety and robustness of LLMs regarding the position of explicit normal instructions, word replacements (verbs in explicit normal instructions, target groups in malicious instructions, cue words for explicit normal instructions), and instruction replacements (different explicit normal instructions). Our results demonstrate that current LLMs not only prioritize certain instruction verbs but also exhibit varying jailbreak rates for different instruction verbs in explicit normal instructions. Code and data are available at https://github.com/qiuhuachuan/latent-jailbreak.

Large language models are becoming increasingly integrated into our lives. Hence, it is important to understand the biases present in their outputs in order to avoid perpetuating harmful stereotypes, which originate in our own flawed ways of thinking. This challenge requires developing new benchmarks and methods for quantifying affective and semantic bias, keeping in mind that LLMs act as psycho-social mirrors that reflect the views and tendencies that are prevalent in society. One such tendency that has harmful negative effects is the global phenomenon of anxiety toward math and STEM subjects. Here, we investigate perceptions of math and STEM fields provided by cutting-edge language models, namely GPT-3, Chat-GPT, and GPT-4, by applying an approach from network science and cognitive psychology. Specifically, we use behavioral forma mentis networks (BFMNs) to understand how these LLMs frame math and STEM disciplines in relation to other concepts. We use data obtained by probing the three LLMs in a language generation task that has previously been applied to humans. Our findings indicate that LLMs have an overall negative perception of math and STEM fields, with math being perceived most negatively. We observe significant differences across the three LLMs. We observe that newer versions (i.e. GPT-4) produce richer, more complex perceptions as well as less negative perceptions compared to older versions and N=159 high-school students. These findings suggest that advances in the architecture of LLMs may lead to increasingly less biased models that could even perhaps someday aid in reducing harmful stereotypes in society rather than perpetuating them.

Semantic networks provide a useful tool to understand how related concepts are retrieved from memory. However, most current network approaches use pairwise links to represent memory recall patterns. Pairwise connections neglect higher-order associations, i.e. relationships between more than two concepts at a time. These higher-order interactions might covariate with (and thus contain information about) how similar concepts are along psycholinguistic dimensions like arousal, valence, familiarity, gender and others. We overcome these limits by introducing feature-rich cognitive hypergraphs as quantitative models of human memory where: (i) concepts recalled together can all engage in hyperlinks involving also more than two concepts at once (cognitive hypergraph aspect), and (ii) each concept is endowed with a vector of psycholinguistic features (feature-rich aspect). We build hypergraphs from word association data and use evaluation methods from machine learning features to predict concept concreteness. Since concepts with similar concreteness tend to cluster together in human memory, we expect to be able to leverage this structure. Using word association data from the Small World of Words dataset, we compared a pairwise network and a hypergraph with N=3586 concepts/nodes. Interpretable artificial intelligence models trained on (1) psycholinguistic features only, (2) pairwise-based feature aggregations, and on (3) hypergraph-based aggregations show significant differences between pairwise and hypergraph links. Specifically, our results show that higher-order and feature-rich hypergraph models contain richer information than pairwise networks leading to improved prediction of word concreteness. The relation with previous studies about conceptual clustering and compartmentalisation in associative knowledge and human memory are discussed.