Yun-Shiuan Chuang, Sameer Narendran, Nikunj Harlalka, Alexander Cheung, Sizhe Gao, Siddharth Suresh, Junjie Hu, Timothy T. Rogers. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

We present a novel multimodal preference dataset for creative tasks, consisting of over 250 million human votes on more than 2.2 million captions, collected through crowdsourcing rating data for The New Yorker's weekly cartoon caption contest over the past eight years. This unique dataset supports the development and evaluation of multimodal large language models and preference-based fine-tuning algorithms for humorous caption generation. We propose novel benchmarks for judging the quality of model-generated captions, utilizing both GPT4 and human judgments to establish ranking-based evaluation strategies. Our experimental results highlight the limitations of current fine-tuning methods, such as RLHF and DPO, when applied to creative tasks. Furthermore, we demonstrate that even state-of-the-art models like GPT4 and Claude currently underperform top human contestants in generating humorous captions. As we conclude this extensive data collection effort, we release the entire preference dataset to the research community, fostering further advancements in AI humor generation and evaluation.

Neural network models of language have long been used as a tool for developing hypotheses about conceptual representation in the mind and brain.For many years, such use involved extracting vector-space representations of words and using distances among these to predict or understand human behavior in various semantic tasks.Contemporary large language models (LLMs), however, make it possible to interrogate the latent structure of conceptual representations using experimental methods nearly identical to those commonly used with human participants.The current work utilizes three common techniques borrowed from cognitive psychology to estimate and compare the structure of concepts in humans and a suite of LLMs.In humans, we show that conceptual structure is robust to differences in culture, language, and method of estimation.Structures estimated from LLM behavior, while individually fairly consistent with those estimated from human behavior, vary much more depending upon the particular task used to generate responsesacross tasks, estimates of conceptual structure from the very same model cohere less with one another than do human structure estimates.These results highlight an important difference between contemporary LLMs and human cognition, with implications for understanding some fundamental limitations of contemporary machine language.

Representational Similarity Learning (RSL) aims to discover features that are important in representing (human-judged) similarities among objects. RSL can be posed as a sparsity-regularized multi-task regression problem. Standard methods, like group lasso, may not select important features if they are strongly correlated with others. To address this shortcoming we present a new regularizer for multitask regression called Group Ordered Weighted \ell_1 (GrOWL). Another key contribution of our paper is a novel application to fMRI brain imaging. Representational Similarity Analysis (RSA) is a tool for testing whether localized brain regions encode perceptual similarities. Using GrOWL, we propose a new approach called Network RSA that can discover arbitrarily structured brain networks (possibly widely distributed and non-local) that encode similarity information. We show, in theory and fMRI experiments, how GrOWL deals with strongly correlated covariates.

Imagine a random walk that outputs a state only when visiting it for the first time. The observed output is therefore a repeat-censored version of the underlying walk, and consists of a permutation of the states or a prefix of it. We call this model initial-visit emitting random walk (INVITE). Prior work has shown that the random walks with such a repeat-censoring mechanism explain well human behavior in memory search tasks, which is of great interest in both the study of human cognition and various clinical applications. However, parameter estimation in INVITE is challenging, because naive likelihood computation by marginalizing over infinitely many hidden random walk trajectories is intractable. In this paper, we propose the first efficient maximum likelihood estimate (MLE) for INVITE by decomposing the censored output into a series of absorbing random walks. We also prove theoretical properties of the MLE including identifiability and consistency. We show that INVITE outperforms several existing methods on real-world human response data from memory search tasks.

Human-generated lists are a form of non-iid data with important applications in machine learning and cognitive psychology. We propose a generative model - sampling with reduced replacement (SWIRL) - for such lists. We discuss SWIRL’s relation to standard sampling paradigms, provide the maximum likelihood estimate for learning, and demonstrate its value with two real-world applications: (i) In a ""feature volunteering"" task where non-experts spontaneously generate feature=>label pairs for text classification, SWIRL improves the accuracy of state-of-the-art feature-learning frameworks. (ii) In a ""verbal fluency"" task where brain-damaged patients generate word lists when prompted with a category, SWIRL parameters align well with existing psychological theories, and our model can classify healthy people vs. patients from the lists they generate.

Multitask learning can be effective when features useful in one task are also useful for other tasks, and the group lasso is a standard method for selecting a common subset of features. In this paper, we are interested in a less restrictive form of multitask learning, wherein (1) the available features can be organized into subsets according to a notion of similarity and (2) features useful in one task are similar, but not necessarily identical, to the features best suited for other tasks. The main contribution of this paper is a new procedure called {\em Sparse Overlapping Sets (SOS) lasso}, a convex optimization that automatically selects similar features for related learning tasks. Error bounds are derived for SOSlasso and its consistency is established for squared error loss. In particular, SOSlasso is motivated by multi-subject fMRI studies in which functional activity is classified using brain voxels as features. Experiments with real and synthetic data demonstrate the advantages of SOSlasso compared to the lasso and group lasso.

An important problem in cognitive psychology is to quantify the perceived similarities between stimuli. Previous work attempted to address this problem with multidimensional scaling (MDS) and its variants. However, there are several shortcomings of the MDS approaches. We propose Yada, a novel general metric-learning procedure based on two-alternative forced-choice behavioral experiments. Our method learns forward and backward nonlinear mappings between an objective space in which the stimuli are defined by the standard feature vector representation and a subjective space in which the distance between a pair of stimuli corresponds to their perceived similarity. We conduct experiments on both synthetic and real human behavioral datasets to assess the effectiveness of Yada. The results show that Yada outperforms several standard embedding and metric-learning algorithms, both in terms of likelihood and recovery error.

We consider the task of human collaborative category learning, where two people work together to classify test items into appropriate categories based on what they learn from a training set. We propose a novel collaboration policy based on the Co-Training algorithm in machine learning, in which the two people play the role of the base learners. The policy restricts each learner's view of the data and limits their communication to only the exchange of their labelings on test items. In a series of empirical studies, we show that the Co-Training policy leads collaborators to jointly produce unique and potentially valuable classification outcomes that are not generated under other collaboration policies. We further demonstrate that these observations can be explained with appropriate machine learning models.

When the distribution of unlabeled data in feature space lies along a manifold, the information it provides may be used by a learner to assist classification in a semi-supervised setting. While manifold learning is well-known in machine learning, the use of manifolds in human learning is largely unstudied. We perform a set of experiments which test a human's ability to use a manifold in a semi-supervised learning task, under varying conditions. We show that humans may be encouraged into using the manifold, overcoming the strong preference for a simple, axis-parallel linear boundary.

We propose to use Rademacher complexity, originally developed in computational learning theory, as a measure of human learning capacity. Rademacher complexity measures a learners ability to fit random data, and can be used to bound the learners true error based on the observed training sample error. We first review the definition of Rademacher complexity and its generalization bound. We then describe a learning the noise" procedure to experimentally measure human Rademacher complexities. The results from empirical studies showed that: (i) human Rademacher complexity can be successfully measured, (ii) the complexity depends on the domain and training sample size in intuitive ways, (iii) human learning respects the generalization bounds, (iv) the bounds can be useful in predicting the danger of overfitting in human learning. Finally, we discuss the potential applications of human Rademacher complexity in cognitive science."

We investigate a topic at the interface of machine learning and cognitive science. Human active learning, where learners can actively query the world for information, is contrasted with passive learning from random examples. Furthermore, we compare human active learning performance with predictions from statistical learning theory. We conduct a series of human category learning experiments inspired by a machine learning task for which active and passive learning error bounds are well understood, and dramatically distinct. Our results indicate that humans are capable of actively selecting informative queries, and in doing so learn better and faster than if they are given random training data, as predicted by learning theory. However, the improvement over passive learning is not as dramatic as that achieved by machine active learning algorithms. To the best of our knowledge, this is the first quantitative study comparing human category learning in active versus passive settings.