For many low-resource languages, the only available language models are large multilingual models trained on many languages simultaneously. Despite state-of-the-art performance on reasoning tasks, we find that these models still struggle with basic grammatical text generation in many languages. First, large multilingual models perform worse than bigrams for many languages (e.g. 24% of languages in XGLM 4.5B; 43% in BLOOM 7.1B) using FLORES perplexity as an evaluation metric. Second, when we train small monolingual models with only 125M parameters on 1GB or less data for 350 languages, these small models outperform large multilingual models both in perplexity and on a massively multilingual grammaticality benchmark. To facilitate future work on low-resource language modeling, we release Goldfish, a suite of over 1,000 small monolingual language models trained comparably for 350 languages. These models represent the first publicly-available monolingual language models for 215 of the languages included.

Language models perform differently across languages. It has been previously suggested that morphological typology may explain some of this variability (Cotterell et al., 2018). We replicate previous analyses and find additional new evidence for a performance gap between agglutinative and fusional languages, where fusional languages, such as English, tend to have better language modeling performance than morphologically more complex languages like Turkish. We then propose and test three possible causes for this performance gap: morphological alignment of tokenizers, tokenization quality, and disparities in dataset sizes and measurement. To test the morphological alignment hypothesis, we present MorphScore, a tokenizer evaluation metric, and supporting datasets for 22 languages. We find some evidence that tokenization quality explains the performance gap, but none for the role of morphological alignment. Instead we find that the performance gap is most reduced when training datasets are of equivalent size across language types, but only when scaled according to the so-called “byte-premium”—the different encoding efficiencies of different languages and orthographies. These results suggest that languages of particular morphological types are not intrinsically advantaged or disadvantaged in language modeling. Differences in performance can be attributed to disparities in dataset size. These findings bear on ongoing efforts to improve performance for low-performing and under-resourced languages.

Can language models reliably predict that possible events are more likely than merely improbable ones? By teasing apart possibility, typicality, and contextual relatedness, we show that despite the results of previous work, language models’ ability to do this is far from robust. In fact, under certain conditions, all models tested—including Llama 3, Gemma 2, and Mistral NeMo—perform at worse-than-chance level, assigning higher probabilities to impossible sentences such as ‘the car was given a parking ticket by the brake’ than to merely unlikely sentences such as ‘the car was given a parking ticket by the explorer’.

Crosslingual transfer is crucial to contemporary language models’ multilingual capabilities, but how it occurs is not well understood. Weask what happens to a monolingual language model when it begins to be trained on a second language. Specifically, we train small bilingual models for which we control the amount of data for each language and the order of language exposure. To find evidence of shared multilingual representations, we turn to structural priming, a method used to study grammatical representations in humans. We first replicate previous crosslingual structural priming results and find that after controlling for training data quantity and language exposure, there are asymmetrical effects across language pairs and directions. We argue that this asymmetry may shape hypotheses about human structural priming effects. We also find that structural priming effects are less robust for less similar language pairs, highlighting potential limitations of crosslingual transfer learning and shared representations for typologically diverse languages.

Large language models (LLMs) have exhibited certain indirect pragmatic capabilities, including interpreting indirect requests and non-literal meanings. Yet, it is unclear whether the success of LLMs on pragmatic tasks generalizes to phenomena that directly probe inferences about the beliefs of others. Indeed, LLMs’ performance on Theory of Mind (ToM) tasks is mixed. To date, the most successful computationally explicit approach to making inferences about others’ beliefs is the Rational Speech Act (RSA) framework, a Bayesian probabilistic model that encodes explicit representations of beliefs. In the present study, we ask whether LLMs outperform RSA in predicting human belief inferences, even though they do not explicitly encode belief representations. We focus specifically on projection inferences, a type of inference that directly probes belief attribution. We find that some LLMs are sensitive to factors that affect the inference process similarly to humans, yet there remains variance in human behavior not fully captured by LLMs. The RSA model, on the other hand, outperforms LLMs in capturing the variances in human data, suggesting that explicit belief representation might be necessary to construct human-like projection inferences.

Everyday AI detection requires differentiating between humans and AI in informal, online conversations. At present, human users most often do not interact directly with bots but instead read their conversations with other humans. We measured how well humans and large language models can discriminate using two modified versions of the Turing test: inverted and displaced. GPT-3.5, GPT-4, and displaced human adjudicators judged whether an agent was human or AI on the basis of a Turing test transcript. We found that both AI and displaced human judges were less accurate than interactive interrogators, with below chance accuracy overall. Moreover, all three judged the best-performing GPT-4 witness to be human more often than human witnesses. This suggests that both humans and current LLMs struggle to distinguish between the two when they are not actively interrogating the person, underscoring an urgent need for more accurate tools to detect AI in conversations.

How should text dataset sizes be compared across languages? Even for content-matched (parallel) corpora, UTF-8 encoded text can require a dramatically different number of bytes for different languages. In our work, we define the byte premium between two languages as the ratio of bytes used to encode content-matched text in those languages. We compute byte premiums for 1155 languages, and we use linear regressions to estimate byte premiums for other languages. We release a tool to obtain byte premiums for any two languages, enabling comparisons of dataset sizes across languages for more equitable multilingual model development and data practices.

We address a growing debate about the extent to which large language models (LLMs) produce behavior consistent with Theory of Mind (ToM) in humans. We present EPITOME: a battery of six experiments that tap diverse ToM capacities, including belief attribution, emotional inference, and pragmatic reasoning. We elicit a performance baseline from human participants for each task. We use the dataset to ask whether distributional linguistic information learned by LLMs is sufficient to explain ToM in humans. We compare performance of five LLMs to a baseline of responses from human comprehenders. Results are mixed. LLMs display considerable sensitivity to mental states and match human performance in several tasks. Yet, they commit systematic errors in others, especially those requiring pragmatic reasoning on the basis of mental state information. Such uneven performance indicates that human-level ToM may require resources beyond distributional information.

A common approach to interpreting multilingual language models is to evaluate their internal representations. For example, studies have found that languages occupy distinct subspaces in the models’ representation spaces, and geometric distances between languages often reflect linguistic properties such as language families and typological features. In our work, we investigate whether geometric distances between language representations correlate with zero-shot crosslingual transfer performance for POS-tagging and NER in three multilingual language models. We consider four distance metrics, including new metrics that identify a basis for a multilingual representation space that sorts axes based on their language-separability. We find that each distance metric either only moderately correlates or does not correlate with crosslingual transfer performance, and metrics do not generalize well across models, layers, and tasks. Although pairwise language separability is a reasonable predictor of crosslingual transfer, representational geometry overall is an inconsistent predictor for the crosslingual performance of multilingual language models.

Transformer language models have received widespread public attention, yet their generated text is often surprising even to NLP researchers. In this survey, we discuss over 250 recent studies of English language model behavior before task-specific fine-tuning. Language models possess basic capabilities in syntax, semantics, pragmatics, world knowledge, and reasoning, but these capabilities are sensitive to specific inputs and surface features. Despite dramatic increases in generated text quality as models scale to hundreds of billions of parameters, the models are still prone to unfactual responses, commonsense errors, memorized text, and social biases. Many of these weaknesses can be framed as over-generalizations or under-generalizations of learned patterns in text. We synthesize recent results to highlight what is currently known about large language model capabilities, thus providing a resource for applied work and for research in adjacent fields that use language models.

Multilingual language models are widely used to extend NLP systems to low-resource languages. However, concrete evidence for the effects of multilinguality on language modeling performance in individual languages remains scarce. Here, we pre-train over 10,000 monolingual and multilingual language models for over 250 languages, including multiple language families that are under-studied in NLP. We assess how language modeling performance in each language varies as a function of (1) monolingual dataset size, (2) added multilingual dataset size, (3) linguistic similarity of the added languages, and (4) model size (up to 45M parameters). We find that in moderation, adding multilingual data improves low-resource language modeling performance, similar to increasing low-resource dataset sizes by up to 33%. Improvements depend on the syntactic similarity of the added multilingual data, with marginal additional effects of vocabulary overlap. However, high-resource languages consistently perform worse in multilingual pre-training scenarios. As dataset sizes increase, adding multilingual data begins to hurt performance for both low-resource and high-resource languages, likely due to limited model capacity (the “curse of multilinguality”). These results suggest that massively multilingual pre-training may not be optimal for any languages involved, but that more targeted models can significantly improve performance.

How do language models learn to make predictions during pre-training? To study this, we extract learning curves from five autoregressive English language model pre-training runs, for 1M unseen tokens in context. We observe that the language models generate short repetitive phrases before learning to generate longer and more coherent text. We also find that individual tokens often exhibit sudden increases or decreases in loss that are surprisingly consistent across pre-training runs. To better understand these fluctuations, we quantify the final surprisal, within-run variability, age of acquisition, forgettability, and cross-run variability of learning curves for individual tokens in context. More frequent tokens reach lower final surprisals, exhibit less variability within and across pre-training runs, are learned earlier, and are less likely to be “forgotten” during pre-training. Higher n-gram probabilities further accentuate these effects. Independent of the target token, shorter and more frequent contexts correlate with marginally more stable and quickly acquired predictions. Based on our results, we argue for the existence of sequential learning dependencies between different model capabilities, and we characterize language model learning as early n-gram learning before gradual refinement of tail n-gram predictions.

We evaluated GPT-4 in a public online Turing test. The best-performing GPT-4 prompt passed in 49.7% of games, outperforming ELIZA (22%) and GPT-3.5 (20%), but falling short of the baseline set by human participants (66%). Participants’ decisions were based mainly on linguistic style (35%) and socioemotional traits (27%), supporting the idea that intelligence, narrowly conceived, is not sufficient to pass the Turing test. Participant knowledge about LLMs and number of games played positively correlated with accuracy in detecting AI, suggesting learning and practice as possible strategies to mitigate deception. Despite known limitations as a test of intelligence, we argue that the Turing test continues to be relevant as an assessment of naturalistic communication and deception. AI models with the ability to masquerade as humans could have widespread societal consequences, and we analyse the effectiveness of different strategies and criteria for judging humanlikeness.

Large Language Models (LLMs) have been criticized for failing to connect linguistic meaning to the world—for failing to solve the “symbol grounding problem.” Multimodal Large Language Models (MLLMs) offer a potential solution to this challenge by combining linguistic representations and processing with other modalities. However, much is still unknown about exactly how and to what degree MLLMs integrate their distinct modalities—and whether the way they do so mirrors the mechanisms believed to underpin grounding in humans. In humans, it has been hypothesized that linguistic meaning is grounded through “embodied simulation,” the activation of sensorimotor and affective representations reflecting described experiences. Across four pre-registered studies, we adapt experimental techniques originally developed to investigate embodied simulation in human comprehenders to ask whether MLLMs are sensitive to sensorimotor features that are implied but not explicit in descriptions of an event. In Experiment 1, we find sensitivity to some features (color and shape) but not others (size, orientation, and volume). In Experiment 2, we identify likely bottlenecks to explain an MLLM’s lack of sensitivity. In Experiment 3, we find that despite sensitivity to implicit sensorimotor features, MLLMs cannot fully account for human behavior on the same task. Finally, in Experiment 4, we compare the psychometric predictive power of different MLLM architectures and find that ViLT, a single-stream architecture, is more predictive of human responses to one sensorimotor feature (shape) than CLIP, a dual-encoder architecture—despite being trained on orders of magnitude less data. These results reveal strengths and limitations in the ability of current MLLMs to integrate language with other modalities, and also shed light on the likely mechanisms underlying human language comprehension.

Does inverse scaling only occur as a function of model size, or can it also occur over the course of training? We carry out an exploratory study investigating whether the performance of language models on specific tasks can decrease (while general performance remains high) during training on the language modeling task. We find 8 tasks on which Pythia 12B (Biderman et al., 2023) shows decreased performance over the course of training. Five of these tasks (TruthfulQA-MC1, TruthfulQA-MC2, Hindsight Neglect, Memo Trap, and Pattern Match Suppression) additionally show a consistent relationship whereby larger language models show a greater decrease in performance the more they are trained, despite showing standard (positive) scaling overall. This highlights the importance of testing performance at all relevant benchmarks any time models are trained on additional data, even if their overall performance improves.

How well do language models deal with quantification? In this study, we focus on ‘few’-type quantifiers, as in ‘few children like toys’, which might pose a particular challenge for language models because the sentence components with out the quantifier are likely to co-occur, and ‘few’-type quantifiers are rare. We present 960 English sentence stimuli from two human neurolinguistic experiments to 22 autoregressive transformer models of differing sizes. Not only do all the models perform poorly on ‘few’-type quantifiers, but overall the larger the model, the worse its performance. This inverse scaling is consistent with previous work suggesting that larger models increasingly reflect online rather than offline human processing, and we argue that the decreasing performance of larger models may challenge uses of language models as the basis for natural language systems.

Abstract grammatical knowledge—of parts of speech and grammatical patterns—is key to the capacity for linguistic generalization in humans. But how abstract is grammatical knowledge in large language models? In the human literature, compelling evidence for grammatical abstraction comes from structural priming. A sentence that shares the same grammatical structure as a preceding sentence is processed and produced more readily. Because confounds exist when using stimuli in a single language, evidence of abstraction is even more compelling from crosslingual structural priming, where use of a syntactic structure in one language primes an analogous structure in another language. We measure crosslingual structural priming in large language models, comparing model behavior to human experimental results from eight crosslingual experiments covering six languages, and four monolingual structural priming experiments in three non-English languages. We find evidence for abstract monolingual and crosslingual grammatical representations in the models that function similarly to those found in humans. These results demonstrate that grammatical representations in multilingual language models are not only similar across languages, but they can causally influence text produced in different languages.

Some languages allow arguments to be omitted in certain contexts. Yet human language comprehenders reliably infer the intended referents of these zero pronouns, in part because they construct expectations about which referents are more likely. We ask whether Neural Language Models also extract the same expectations. We test whether 12 contemporary language models display expectations that reflect human behavior when exposed to sentences with zero pronouns from five behavioral experiments conducted in Italian by Carminati (2005). We find that three models - XGLM 2.9B, 4.5B, and 7.5B - capture the human behavior from all the experiments, with others successfully modeling some of the results. This result suggests that human expectations about coreference can be derived from exposure to language, and also indicates features of language models that allow them to better reflect human behavior.

We investigate how neural language models acquire individual words during training, extracting learning curves and ages of acquisition for over 600 words on the MacArthur-Bates Communicative Development Inventory (Fenson et al., 2007). Drawing on studies of word acquisition in children, we evaluate multiple predictors for words’ ages of acquisition in LSTMs, BERT, and GPT-2. We find that the effects of concreteness, word length, and lexical class are pointedly different in children and language models, reinforcing the importance of interaction and sensorimotor experience in child language acquisition. Language models rely far more on word frequency than children, but, like children, they exhibit slower learning of words in longer utterances. Interestingly, models follow consistent patterns during training for both unidirectional and bidirectional models, and for both LSTM and Transformer architectures. Models predict based on unigram token frequencies early in training, before transitioning loosely to bigram probabilities, eventually converging on more nuanced predictions. These results shed light on the role of distributional learning mechanisms in children, while also providing insights for more human-like language acquisition in language models.

Are the predictions of humans and language models affected by similar things? Research suggests that while comprehending language, humans make predictions about upcoming words, with more predictable words being processed more easily. However, evidence also shows that humans display a similar processing advantage for highly anomalous words when these words are semantically related to the preceding context or to the most probable continuation. Using stimuli from 3 psycholinguistic experiments, we find that this is also almost always also the case for 8 contemporary transformer language models (BERT, ALBERT, RoBERTa, XLM-R, GPT-2, GPT-Neo, GPT-J, and XGLM). We then discuss the implications of this phenomenon for our understanding of both human language comprehension and the predictions made by language models.

We assess how multilingual language models maintain a shared multilingual representation space while still encoding language-sensitive information in each language. Using XLM-R as a case study, we show that languages occupy similar linear subspaces after mean-centering, evaluated based on causal effects on language modeling performance and direct comparisons between subspaces for 88 languages. The subspace means differ along language-sensitive axes that are relatively stable throughout middle layers, and these axes encode information such as token vocabularies. Shifting representations by language means is sufficient to induce token predictions in different languages. However, we also identify stable language-neutral axes that encode information such as token positions and part-of-speech. We visualize representations projected onto language-sensitive and language-neutral axes, identifying language family and part-of-speech clusters, along with spirals, toruses, and curves representing token position information. These results demonstrate that multilingual language models encode information along orthogonal language-sensitive and language-neutral axes, allowing the models to extract a variety of features for downstream tasks and cross-lingual transfer learning.

Most words are ambiguous—-i.e., they convey distinct meanings in different contexts—-and even the meanings of unambiguous words are context-dependent. Both phenomena present a challenge for NLP. Recently, the advent of contextualized word embeddings has led to success on tasks involving lexical ambiguity, such as Word Sense Disambiguation. However, there are few tasks that directly evaluate how well these contextualized embeddings accommodate the more continuous, dynamic nature of word meaning—-particularly in a way that matches human intuitions. We introduce RAW-C, a dataset of graded, human relatedness judgments for 112 ambiguous words in context (with 672 sentence pairs total), as well as human estimates of sense dominance. The average inter-annotator agreement (assessed using a leave-one-annotator-out method) was 0.79. We then show that a measure of cosine distance, computed using contextualized embeddings from BERT and ELMo, correlates with human judgments, but that cosine distance also systematically underestimates how similar humans find uses of the same sense of a word to be, and systematically overestimates how similar humans find uses of different-sense homonyms. Finally, we propose a synthesis between psycholinguistic theories of the mental lexicon and computational models of lexical semantics.

We investigate the extent to which word surprisal can be used to predict a neural measure of human language processing difficulty—the N400. To do this, we use recurrent neural networks to calculate the surprisal of stimuli from previously published neurolinguistic studies of the N400. We find that surprisal can predict N400 amplitude in a wide range of cases, and the cases where it cannot do so provide valuable insight into the neurocognitive processes underlying the response.