Polina Tsvilodub

Polina Tsvilodub, Michael Franke, Fausto Carcassi

To what extent can LLMs be used as part of a cognitive model of language generation? In this paper, we approach this question by exploring a neuro-symbolic implementation of an algorithmic cognitive model of referential expression generation by Dale & Reiter (1995). The symbolic task analysis implements the generation as an iterative procedure that scaffolds symbolic and gpt-3.5-turbo-based modules. We compare this implementation to an ablated model and a one-shot LLM-only baseline on the A3DS dataset (Tsvilodub & Franke, 2023). We find that our hybrid approach is cognitively plausible and performs well in complex contexts, while allowing for more open-ended modeling of language generation in a larger domain.

Polina Tsvilodub, Michael Franke

Evaluating grounded neural language model performance with respect to pragmatic qualities like the trade off between truthfulness, contrastivity and overinformativity of generated utterances remains a challenge in absence of data collected from humans. To enable such evaluation, we present a novel open source image-text dataset "Annotated 3D Shapes" (A3DS) comprising over nine million exhaustive natural language annotations and over 12 million variable-granularity captions for the 480,000 images provided by Burges & Kim (2018). We showcase the evaluation of pragmatic abilities developed by a task-neutral image captioner fine-tuned in a multi-agent communication setting to produce contrastive captions. The evaluation is enabled by the dataset because the exhaustive annotations allow to quantify the presence of contrastive features in the model's generations. We show that the model develops human-like patterns (informativity, brevity, over-informativity for specific features (e.g., shape, color biases)).

Marcel Binz, Elif Akata, Abdullah Almaatouq et al.

Large language models (LLMs) are increasingly used as surrogates for human participants, but it remains unclear which models best capture human behavior and why. To address this, we introduce Psych-201, a novel dataset that enables us to measure behavioral alignment at scale. We find that post-training -- the stage that turns base models into useful assistants -- consistently reduces alignment with human behavior across model families, sizes, and objectives. Moreover, this misalignment widens in newer model generations even as base models continue to improve. Finally, we find that persona-induction -- a popular technique for eliciting human-like behavior by conditioning models on participant-specific information -- does not improve predictions at the level of individuals. Taken together, our results suggest that the very processes that are currently employed to turn LLMs into useful assistants also make them less accurate models of human behavior.

Polina Tsvilodub, Hening Wang, Sharon Grosch et al.

This paper systematically compares different methods of deriving item-level predictions of language models for multiple-choice tasks. It compares scoring methods for answer options based on free generation of responses, various probability-based scores, a Likert-scale style rating method, and embedding similarity. In a case study on pragmatic language interpretation, we find that LLM predictions are not robust under variation of method choice, both within a single LLM and across different LLMs. As this variability entails pronounced researcher degrees of freedom in reporting results, knowledge of the variability is crucial to secure robustness of results and research integrity.

Polina Tsvilodub, Kanishk Gandhi, Haoran Zhao et al.

Humans naturally interpret numbers non-literally, effortlessly combining context, world knowledge, and speaker intent. We investigate whether large language models (LLMs) interpret numbers similarly, focusing on hyperbole and pragmatic halo effects. Through systematic comparison with human data and computational models of pragmatic reasoning, we find that LLMs diverge from human interpretation in striking ways. By decomposing pragmatic reasoning into testable components, grounded in the Rational Speech Act framework, we pinpoint where LLM processing diverges from human cognition -- not in prior knowledge, but in reasoning with it. This insight leads us to develop a targeted solution -- chain-of-thought prompting inspired by an RSA model makes LLMs' interpretations more human-like. Our work demonstrates how computational cognitive models can both diagnose AI-human differences and guide development of more human-like language understanding capabilities.

Polina Tsvilodub, Karl Mulligan, Todd Snider et al.

When deciding how to act under uncertainty, agents may choose to act to reduce uncertainty or they may act despite that uncertainty.In communicative settings, an important way of reducing uncertainty is by asking clarification questions (CQs). We predict that the decision to ask a CQ depends on both contextual uncertainty and the cost of alternative actions, and that these factors interact: uncertainty should matter most when acting incorrectly is costly. We formalize this interaction in a computational model based on expected regret: how much an agent stands to lose by acting now rather than with full information. We test these predictions in two experiments, one examining purely linguistic responses to questions and another extending to choices between clarification and non-linguistic action. Taken together, our results suggest a rational tradeoff: humans tend to seek clarification proportional to the risk of substantial loss when acting under uncertainty.

Polina Tsvilodub, Jan-Felix Klumpp, Amir Mohammadpour et al.

This paper investigates whether LMs recruit shared computational mechanisms for general Theory of Mind (ToM) and language-specific pragmatic reasoning in order to contribute to the general question of whether LMs may be said to have emergent "social world models", i.e., representations of mental states that are repurposed across tasks (the functional integration hypothesis). Using behavioral evaluations and causal-mechanistic experiments via functional localization methods inspired by cognitive neuroscience, we analyze LMs' performance across seven subcategories of ToM abilities (Beaudoin et al., 2020) on a substantially larger localizer dataset than used in prior like-minded work. Results from stringent hypothesis-driven statistical testing offer suggestive evidence for the functional integration hypothesis, indicating that LMs may develop interconnected "social world models" rather than isolated competencies. This work contributes novel ToM localizer data, methodological refinements to functional localization techniques, and empirical insights into the emergence of social cognition in artificial systems.

Polina Tsvilodub, Robert D. Hawkins, Michael Franke

Computational models of pragmatic language use have traditionally relied on hand-specified sets of utterances and meanings, limiting their applicability to real-world language use. We propose a neuro-symbolic framework that enhances probabilistic cognitive models by integrating LLM-based modules to propose and evaluate key components in natural language, eliminating the need for manual specification. Through a classic case study of pragmatic question-answering, we systematically examine various approaches to incorporating neural modules into the cognitive model -- from evaluating utilities and literal semantics to generating alternative utterances and goals. We find that hybrid models can match or exceed the performance of traditional probabilistic models in predicting human answer patterns. However, the success of the neuro-symbolic model depends critically on how LLMs are integrated: while they are particularly effective for proposing alternatives and transforming abstract goals into utilities, they face challenges with truth-conditional semantic evaluation. This work charts a path toward more flexible and scalable models of pragmatic language use while illuminating crucial design considerations for balancing neural and symbolic components.

Michael Franke, Polina Tsvilodub, Fausto Carcassi

State of the art large language models (LLMs) have shown impressive performance on a variety of benchmark tasks and are increasingly used as components in larger applications, where LLM-based predictions serve as proxies for human judgements or decision. This raises questions about the human-likeness of LLM-derived information, alignment with human intuition, and whether LLMs could possibly be considered (parts of) explanatory models of (aspects of) human cognition or language use. To shed more light on these issues, we here investigate the human-likeness of LLMs' predictions for multiple-choice decision tasks from the perspective of Bayesian statistical modeling. Using human data from a forced-choice experiment on pragmatic language use, we find that LLMs do not capture the variance in the human data at the item-level. We suggest different ways of deriving full distributional predictions from LLMs for aggregate, condition-level data, and find that some, but not all ways of obtaining condition-level predictions yield adequate fits to human data. These results suggests that assessment of LLM performance depends strongly on seemingly subtle choices in methodology, and that LLMs are at best predictors of human behavior at the aggregate, condition-level, for which they are, however, not designed to, or usually used to, make predictions in the first place.

Polina Tsvilodub, Paul Marty, Sonia Ramotowska et al.

Human communication is based on a variety of inferences that we draw from sentences, often going beyond what is literally said. While there is wide agreement on the basic distinction between entailment, implicature, and presupposition, the status of many inferences remains controversial. In this paper, we focus on three inferences of plain and embedded disjunctions, and compare them with regular scalar implicatures. We investigate this comparison from the novel perspective of the predictions of state-of-the-art large language models, using the same experimental paradigms as recent studies investigating the same inferences with humans. The results of our best performing models mostly align with those of humans, both in the large differences we find between those inferences and implicatures, as well as in fine-grained distinctions among different aspects of those inferences.

Polina Tsvilodub, Michael Franke, Robert D. Hawkins et al.

When faced with a polar question, speakers often provide overinformative answers going beyond a simple "yes" or "no". But what principles guide the selection of additional information? In this paper, we provide experimental evidence from two studies suggesting that overinformativeness in human answering is driven by considerations of relevance to the questioner's goals which they flexibly adjust given the functional context in which the question is uttered. We take these human results as a strong benchmark for investigating question-answering performance in state-of-the-art neural language models, conducting an extensive evaluation on items from human experiments. We find that most models fail to adjust their answering behavior in a human-like way and tend to include irrelevant information. We show that GPT-3 is highly sensitive to the form of the prompt and only achieves human-like answer patterns when guided by an example and cognitively-motivated explanation.