Tal Linzen

Aarohi Srivastava, Abhinav Rastogi, Abhishek Rao et al. · allen-ai, amazon-science

Language models demonstrate both quantitative improvement and new qualitative capabilities with increasing scale. Despite their potentially transformative impact, these new capabilities are as yet poorly characterized. In order to inform future research, prepare for disruptive new model capabilities, and ameliorate socially harmful effects, it is vital that we understand the present and near-future capabilities and limitations of language models. To address this challenge, we introduce the Beyond the Imitation Game benchmark (BIG-bench). BIG-bench currently consists of 204 tasks, contributed by 450 authors across 132 institutions. Task topics are diverse, drawing problems from linguistics, childhood development, math, common-sense reasoning, biology, physics, social bias, software development, and beyond. BIG-bench focuses on tasks that are believed to be beyond the capabilities of current language models. We evaluate the behavior of OpenAI's GPT models, Google-internal dense transformer architectures, and Switch-style sparse transformers on BIG-bench, across model sizes spanning millions to hundreds of billions of parameters. In addition, a team of human expert raters performed all tasks in order to provide a strong baseline. Findings include: model performance and calibration both improve with scale, but are poor in absolute terms (and when compared with rater performance); performance is remarkably similar across model classes, though with benefits from sparsity; tasks that improve gradually and predictably commonly involve a large knowledge or memorization component, whereas tasks that exhibit "breakthrough" behavior at a critical scale often involve multiple steps or components, or brittle metrics; social bias typically increases with scale in settings with ambiguous context, but this can be improved with prompting.

Jackson Petty, Sjoerd van Steenkiste, Ishita Dasgupta et al. · deepmind

To process novel sentences, language models (LMs) must generalize compositionally -- combine familiar elements in new ways. What aspects of a model's structure promote compositional generalization? Focusing on transformers, we test the hypothesis, motivated by theoretical and empirical work, that deeper transformers generalize more compositionally. Simply adding layers increases the total number of parameters; to address this confound between depth and size, we construct three classes of models which trade off depth for width such that the total number of parameters is kept constant (41M, 134M and 374M parameters). We pretrain all models as LMs and fine-tune them on tasks that test for compositional generalization. We report three main conclusions: (1) after fine-tuning, deeper models generalize more compositionally than shallower models do, but the benefit of additional layers diminishes rapidly; (2) within each family, deeper models show better language modeling performance, but returns are similarly diminishing; (3) the benefits of depth for compositional generalization cannot be attributed solely to better performance on language modeling. Because model latency is approximately linear in the number of layers, these results lead us to the recommendation that, with a given total parameter budget, transformers can be made shallower than is typical without sacrificing performance.

Najoung Kim, Tal Linzen, Paul Smolensky

Human linguistic capacity is often characterized by compositionality and the generalization it enables -- human learners can produce and comprehend novel complex expressions by composing known parts. Several benchmarks exploit distributional control across training and test to gauge compositional generalization, where certain lexical items only occur in limited contexts during training. While recent work using these benchmarks suggests that pretrained models achieve impressive generalization performance, we argue that exposure to pretraining data may break the aforementioned distributional control. Using the COGS benchmark of Kim and Linzen (2020), we test two modified evaluation setups that control for this issue: (1) substituting context-controlled lexical items with novel character sequences, and (2) substituting them with special tokens represented by novel embeddings. We find that both of these setups lead to lower generalization performance in T5 (Raffel et al., 2020), suggesting that previously reported results have been overestimated due to uncontrolled lexical exposure during pretraining. The performance degradation is more extreme with novel embeddings, and the degradation increases with the amount of pretraining data, highlighting an interesting case of inverse scaling.

Suhas Arehalli, Brian Dillon, Tal Linzen

Humans exhibit garden path effects: When reading sentences that are temporarily structurally ambiguous, they slow down when the structure is disambiguated in favor of the less preferred alternative. Surprisal theory (Hale, 2001; Levy, 2008), a prominent explanation of this finding, proposes that these slowdowns are due to the unpredictability of each of the words that occur in these sentences. Challenging this hypothesis, van Schijndel & Linzen (2021) find that estimates of the cost of word predictability derived from language models severely underestimate the magnitude of human garden path effects. In this work, we consider whether this underestimation is due to the fact that humans weight syntactic factors in their predictions more highly than language models do. We propose a method for estimating syntactic predictability from a language model, allowing us to weigh the cost of lexical and syntactic predictability independently. We find that treating syntactic predictability independently from lexical predictability indeed results in larger estimates of garden path. At the same time, even when syntactic predictability is independently weighted, surprisal still greatly underestimate the magnitude of human garden path effects. Our results support the hypothesis that predictability is not the only factor responsible for the processing cost associated with garden path sentences.

Matthew Mandelkern, Tal Linzen

What do language models (LMs) do with language? Everyone agrees that they can produce sequences of (mostly) coherent strings of English. But do those sentences mean something, or are LMs simply babbling in a convincing simulacrum of language use? Here we will address one aspect of this broad question: whether LMs' words can refer, that is, achieve "word-to-world" connections. There is prima facie reason to think they do not since LMs do not interact with the world in the way that ordinary language users do. Drawing on insights from the externalist tradition in philosophy of language, we argue that those appearances are misleading: even if the inputs to an LM are simply strings of text, they are strings of text with natural histories, and that may suffice to put LMs' words into referential contact with the external world.

Aaron Mueller, Robert Frank, Tal Linzen et al.

Relations between words are governed by hierarchical structure rather than linear ordering. Sequence-to-sequence (seq2seq) models, despite their success in downstream NLP applications, often fail to generalize in a hierarchy-sensitive manner when performing syntactic transformations - for example, transforming declarative sentences into questions. However, syntactic evaluations of seq2seq models have only observed models that were not pre-trained on natural language data before being trained to perform syntactic transformations, in spite of the fact that pre-training has been found to induce hierarchical linguistic generalizations in language models; in other words, the syntactic capabilities of seq2seq models may have been greatly understated. We address this gap using the pre-trained seq2seq models T5 and BART, as well as their multilingual variants mT5 and mBART. We evaluate whether they generalize hierarchically on two transformations in two languages: question formation and passivization in English and German. We find that pre-trained seq2seq models generalize hierarchically when performing syntactic transformations, whereas models trained from scratch on syntactic transformations do not. This result presents evidence for the learnability of hierarchical syntactic information from non-annotated natural language text while also demonstrating that seq2seq models are capable of syntactic generalization, though only after exposure to much more language data than human learners receive.

Sebastian Schuster, Tal Linzen

Understanding longer narratives or participating in conversations requires tracking of discourse entities that have been mentioned. Indefinite noun phrases (NPs), such as 'a dog', frequently introduce discourse entities but this behavior is modulated by sentential operators such as negation. For example, 'a dog' in 'Arthur doesn't own a dog' does not introduce a discourse entity due to the presence of negation. In this work, we adapt the psycholinguistic assessment of language models paradigm to higher-level linguistic phenomena and introduce an English evaluation suite that targets the knowledge of the interactions between sentential operators and indefinite NPs. We use this evaluation suite for a fine-grained investigation of the entity tracking abilities of the Transformer-based models GPT-2 and GPT-3. We find that while the models are to a certain extent sensitive to the interactions we investigate, they are all challenged by the presence of multiple NPs and their behavior is not systematic, which suggests that even models at the scale of GPT-3 do not fully acquire basic entity tracking abilities.

Aditya Yedetore, Tal Linzen, Robert Frank et al.

When acquiring syntax, children consistently choose hierarchical rules over competing non-hierarchical possibilities. Is this preference due to a learning bias for hierarchical structure, or due to more general biases that interact with hierarchical cues in children's linguistic input? We explore these possibilities by training LSTMs and Transformers - two types of neural networks without a hierarchical bias - on data similar in quantity and content to children's linguistic input: text from the CHILDES corpus. We then evaluate what these models have learned about English yes/no questions, a phenomenon for which hierarchical structure is crucial. We find that, though they perform well at capturing the surface statistics of child-directed speech (as measured by perplexity), both model types generalize in a way more consistent with an incorrect linear rule than the correct hierarchical rule. These results suggest that human-like generalization from text alone requires stronger biases than the general sequence-processing biases of standard neural network architectures.

William Merrill, Alex Warstadt, Tal Linzen

Language models are often trained on text alone, without additional grounding. There is debate as to how much of natural language semantics can be inferred from such a procedure. We prove that entailment judgments between sentences can be extracted from an ideal language model that has perfectly learned its target distribution, assuming the training sentences are generated by Gricean agents, i.e., agents who follow fundamental principles of communication from the linguistic theory of pragmatics. We also show entailment judgments can be decoded from the predictions of a language model trained on such Gricean data. Our results reveal a pathway for understanding the semantic information encoded in unlabeled linguistic data and a potential framework for extracting semantics from language models.

Aaron Mueller, Yu Xia, Tal Linzen

Structural probing work has found evidence for latent syntactic information in pre-trained language models. However, much of this analysis has focused on monolingual models, and analyses of multilingual models have employed correlational methods that are confounded by the choice of probing tasks. In this study, we causally probe multilingual language models (XGLM and multilingual BERT) as well as monolingual BERT-based models across various languages; we do this by performing counterfactual perturbations on neuron activations and observing the effect on models' subject-verb agreement probabilities. We observe where in the model and to what extent syntactic agreement is encoded in each language. We find significant neuron overlap across languages in autoregressive multilingual language models, but not masked language models. We also find two distinct layer-wise effect patterns and two distinct sets of neurons used for syntactic agreement, depending on whether the subject and verb are separated by other tokens. Finally, we find that behavioral analyses of language models are likely underestimating how sensitive masked language models are to syntactic information.

Kristijan Armeni, Christopher Honey, Tal Linzen

When a language model is trained to predict natural language sequences, its prediction at each moment depends on a representation of prior context. What kind of information about the prior context can language models retrieve? We tested whether language models could retrieve the exact words that occurred previously in a text. In our paradigm, language models (transformers and an LSTM) processed English text in which a list of nouns occurred twice. We operationalized retrieval as the reduction in surprisal from the first to the second list. We found that the transformers retrieved both the identity and ordering of nouns from the first list. Further, the transformers' retrieval was markedly enhanced when they were trained on a larger corpus and with greater model depth. Lastly, their ability to index prior tokens was dependent on learned attention patterns. In contrast, the LSTM exhibited less precise retrieval, which was limited to list-initial tokens and to short intervening texts. The LSTM's retrieval was not sensitive to the order of nouns and it improved when the list was semantically coherent. We conclude that transformers implemented something akin to a working memory system that could flexibly retrieve individual token representations across arbitrary delays; conversely, the LSTM maintained a coarser and more rapidly-decaying semantic gist of prior tokens, weighted toward the earliest items.

Cara Su-Yi Leong, Tal Linzen

Grammatical rules in natural languages are often characterized by exceptions. How do language learners learn these exceptions to otherwise general patterns? Here, we study this question through the case study of English passivization. While passivization is in general quite productive, there are cases where it cannot apply (cf. the following sentence is ungrammatical: *One hour was lasted by the meeting). Using neural network language models as theories of language acquisition, we explore the sources of indirect evidence that a learner can leverage to learn whether a verb can be passivized. We first characterize English speakers' judgments of exceptions to the passive, and confirm that speakers find some verbs more passivizable than others. We then show that a neural network language model's verb passivizability judgments are largely similar to those displayed by humans, suggesting that evidence for these exceptions is available in the linguistic input. Finally, we test two hypotheses as to the source of evidence that language models use to learn these restrictions: frequency (entrenchment) and semantics (affectedness). We do so by training models on versions of the corpus that have had sentences of the types implicated by each hypothesis removed, altered, or introduced. We find support for both hypotheses: entrenchment and affectedness make independent contributions to a verb's passivizability. From a methodological point of view, this study highlights the utility of altering a language model's training data for answering questions where complete control over a learner's input is vital.

Leshem Choshen, Ryan Cotterell, Mustafa Omer Gul et al. · ibm-research

The goal of the BabyLM is to stimulate new research connections between cognitive modeling and language model pretraining. We invite contributions in this vein to the BabyLM Workshop, which will also include the 4th iteration of the BabyLM Challenge. As in previous years, the challenge features two ``standard'' tracks (Strict and Strict-Small), in which participants must train language models on under 100M or 10M words of data, respectively. This year, we move beyond our previous English-only pretraining datasets with a new Multilingual track, focusing on English, Dutch, and Chinese. For the workshop, we call for papers related to the overall theme of BabyLM, which includes training efficiency, small-scale training datasets, cognitive modeling, model evaluation, and architecture innovation.

Cara Su-Yi Leong, Tal Linzen

Artificial neural networks can generalize productively to novel contexts. Can they also learn exceptions to those productive rules? We explore this question using the case of restrictions on English passivization (e.g., the fact that "The vacation lasted five days" is grammatical, but "*Five days was lasted by the vacation" is not). We collect human acceptability judgments for passive sentences with a range of verbs, and show that the probability distribution defined by GPT-2, a language model, matches the human judgments with high correlation. We also show that the relative acceptability of a verb in the active vs. passive voice is positively correlated with the relative frequency of its occurrence in those voices. These results provide preliminary support for the entrenchment hypothesis, according to which learners track and uses the distributional properties of their input to learn negative exceptions to rules. At the same time, this hypothesis fails to explain the magnitude of unpassivizability demonstrated by certain individual verbs, suggesting that other cues to exceptionality are available in the linguistic input.

Bingzhi Li, Lucia Donatelli, Alexander Koller et al.

The goal of compositional generalization benchmarks is to evaluate how well models generalize to new complex linguistic expressions. Existing benchmarks often focus on lexical generalization, the interpretation of novel lexical items in syntactic structures familiar from training; structural generalization tasks, where a model needs to interpret syntactic structures that are themselves unfamiliar from training, are often underrepresented, resulting in overly optimistic perceptions of how well models can generalize. We introduce SLOG, a semantic parsing dataset that extends COGS (Kim and Linzen, 2020) with 17 structural generalization cases. In our experiments, the generalization accuracy of Transformer models, including pretrained ones, only reaches 40.6%, while a structure-aware parser only achieves 70.8%. These results are far from the near-perfect accuracy existing models achieve on COGS, demonstrating the role of SLOG in foregrounding the large discrepancy between models' lexical and structural generalization capacities.

Jackson Petty, Sjoerd van Steenkiste, Tal Linzen

Large language models are increasingly trained on corpora containing both natural language and non-linguistic data like source code. Aside from aiding programming-related tasks, anecdotal evidence suggests that including code in pretraining corpora may improve performance on other, unrelated tasks, yet to date no work has been able to establish a causal connection by controlling between language and code data. Here we do just this. We pretrain language models on datasets which interleave natural language and code in two different settings: additive, in which the total volume of data seen during pretraining is held constant; and competitive, in which the volume of language data is held constant. We study how the pretraining mixture affects performance on (a) a diverse collection of tasks included in the BigBench benchmark, and (b) compositionality, measured by generalization accuracy on semantic parsing and syntactic transformations. We find that pretraining on higher proportions of code improves performance on compositional tasks involving structured output (like semantic parsing), and mathematics. Conversely, increase code mixture can harm performance on other tasks, including on tasks that requires sensitivity to linguistic structure such as syntax or morphology, and tasks measuring real-world knowledge.

Tiwalayo Eisape, MH Tessler, Ishita Dasgupta et al.

A central component of rational behavior is logical inference: the process of determining which conclusions follow from a set of premises. Psychologists have documented several ways in which humans' inferences deviate from the rules of logic. Do language models, which are trained on text generated by humans, replicate such human biases, or are they able to overcome them? Focusing on the case of syllogisms -- inferences from two simple premises -- we show that, within the PaLM2 family of transformer language models, larger models are more logical than smaller ones, and also more logical than humans. At the same time, even the largest models make systematic errors, some of which mirror human reasoning biases: they show sensitivity to the (irrelevant) ordering of the variables in the syllogism, and draw confident but incorrect inferences from particular syllogisms (syllogistic fallacies). Overall, we find that language models often mimic the human biases included in their training data, but are able to overcome them in some cases.

William Timkey, Tal Linzen

Two of the central factors believed to underpin human sentence processing difficulty are expectations and retrieval from working memory. A recent attempt to create a unified cognitive model integrating these two factors relied on the parallels between the self-attention mechanism of transformer language models and cue-based retrieval theories of working memory in human sentence processing (Ryu and Lewis 2021). While Ryu and Lewis show that attention patterns in specialized attention heads of GPT-2 are consistent with similarity-based interference, a key prediction of cue-based retrieval models, their method requires identifying syntactically specialized attention heads, and makes the cognitively implausible assumption that hundreds of memory retrieval operations take place in parallel. In the present work, we develop a recurrent neural language model with a single self-attention head, which more closely parallels the memory system assumed by cognitive theories. We show that our model's single attention head captures semantic and syntactic interference effects observed in human experiments.

Aaron Mueller, Albert Webson, Jackson Petty et al.

In-context learning (ICL) is now a common method for teaching large language models (LLMs) new tasks: given labeled examples in the input context, the LLM learns to perform the task without weight updates. Do models guided via ICL infer the underlying structure of the task defined by the context, or do they rely on superficial heuristics that only generalize to identically distributed examples? We address this question using transformations tasks and an NLI task that assess sensitivity to syntax - a requirement for robust language understanding. We further investigate whether out-of-distribution generalization can be improved via chain-of-thought prompting, where the model is provided with a sequence of intermediate computation steps that illustrate how the task ought to be performed. In experiments with models from the GPT, PaLM, and Llama 2 families, we find large variance across LMs. The variance is explained more by the composition of the pre-training corpus and supervision methods than by model size; in particular, models pre-trained on code generalize better, and benefit more from chain-of-thought prompting.

Dario Paape, Tal Linzen, Shravan Vasishth

Using temporarily ambiguous garden-path sentences ("While the team trained the striker wondered ...") as a test case, we present a latent-process mixture model of human reading behavior across four different reading paradigms (eye tracking, uni- and bidirectional self-paced reading, Maze). The model distinguishes between garden-path probability, garden-path cost, and reanalysis cost, and yields more realistic processing cost estimates by taking into account trials with inattentive reading. We show that the model is able to reproduce empirical patterns with regard to rereading behavior, comprehension question responses, and grammaticality judgments. Cross-validation reveals that the mixture model also has better predictive fit to human reading patterns and end-of-trial task data than a mixture-free model based on GPT-2-derived surprisal values. We discuss implications for future work.

Shashwat Singh, Tal Linzen, Shauli Ravfogel

Can large language models detect and report their own internal states? A number of studies have argued that the answer to this question is yes. We argue, based on lessons from human metacognition research, that this conclusion may be premature: to be convinced of this conclusion we need to distinguish genuine introspection from pattern matching based on surface-level cues. Furthermore, we argue that behavioral evidence alone is inherently insufficient to establish strong introspective claims. We re-examine two recently introduced evaluation paradigms in light of this consideration. In the first paradigm, models are expected to detect whether their internal states have been tampered with. We find that models cannot reliably distinguish such interventions on their internal states from manipulations of the input, suggesting that their success in the original studies reflects their ability to detect anomalies more generally, as opposed to interventions on their internal states in particular. In the second paradigm we examine, models are tasked with predicting labels derived from their own hidden states. Here, we find that classifiers that only have access to the input achieve equivalent performance to the model's own in-context predictions, indicating that the original results do not conclusively demonstrate that the model has privileged access to its internal representations. We further introduce a relabeled control setting, where models cannot rely on the semantics of the task to solve it, and instead must rely on the internal representation; models perform closer to chance on this better-controlled version of the task. Taken together, these results indicate that current evidence is insufficient to establish that LLMs display metacognitive monitoring.

Qihan Wang, Nicholas Tomlin, Michael Hu et al.

Language models are increasingly being deployed as user simulators, but their memory is far more reliable than that of real users. To measure this gap, we run a series of classic memory experiments from psychology on both humans and language models. Across tasks, we find that out-of-the-box language models exhibit better memory than humans, even when prompted to imitate human behavior. We then show that better prompting strategies and the use of a compactor can cause language models to forget content in a more human-like way. Using these methods, we show preliminary evidence that language models with human-like memory constraints can function as more effective user simulators in a downstream education task. Finally, we release human reference data and benchmarks to support future work on simulating human memory with language models.

Sophie Hao, Tal Linzen

Deep architectures such as Transformers are sometimes criticized for having uninterpretable "black-box" representations. We use causal intervention analysis to show that, in fact, some linguistic features are represented in a linear, interpretable format. Specifically, we show that BERT's ability to conjugate verbs relies on a linear encoding of subject number that can be manipulated with predictable effects on conjugation accuracy. This encoding is found in the subject position at the first layer and the verb position at the last layer, but distributed across positions at middle layers, particularly when there are multiple cues to subject number.

Michael A. Lepori, Tal Linzen, Ann Yuan et al.

Though large language models (LLMs) have enabled great success across a wide variety of tasks, they still appear to fall short of one of the loftier goals of artificial intelligence research: creating an artificial system that can adapt its behavior to radically new contexts upon deployment. One important step towards this goal is to create systems that can induce rich representations of data that are seen in-context, and then flexibly deploy these representations to accomplish goals. Recently, Park et al. (2024) demonstrated that current LLMs are indeed capable of inducing such representation from context (i.e., in-context representation learning). The present study investigates whether LLMs can use these representations to complete simple downstream tasks. We first assess whether open-weights LLMs can use in-context representations for next-token prediction, and then probe models using a novel task, adaptive world modeling. In both tasks, we find evidence that open-weights LLMs struggle to deploy representations of novel semantics that are defined in-context, even if they encode these semantics in their latent representations. Furthermore, we assess closed-source, state-of-the-art reasoning models on the adaptive world modeling task, demonstrating that even the most performant LLMs cannot reliably leverage novel patterns presented in-context. Overall, this work seeks to inspire novel methods for encouraging models to not only encode information presented in-context, but to do so in a manner that supports flexible deployment of this information.

William Timkey, Brian Dillon, Tal Linzen

Surprisal theory posits that the processing difficulty of a word is determined by its predictability in context, offering a potential link between human sentence processing and next-word predictions from language models. While language model (LM) surprisals successfully predict reading times in naturalistic text, they systematically underpredict the magnitude of difficulty observed in controlled studies of syntactic ambiguity, particularly in garden path sentences. This mismatch might arise from differences in the computational constraints between humans and LMs. Here we test one such hypothesis, specifically, that LMs may be able to simultaneously consider a greater number of distinct sentence interpretations at once, compared to humans. Using Recurrent Neural Network Grammars (RNNGs) with word-synchronous beam search, we systematically vary the number of simultaneous parses used to compute word surprisal, and then use these surprisals to predict human reading times. Reducing the number of simultaneous active parses indeed increases the magnitude of predicted garden path effects, but not nearly enough to capture the full magnitude of the effects in humans. This suggests that differences in the number of simultaneous parses available to LMs and humans cannot reconcile LM-based surprisal with human sentence processing.

Michael Y. Hu, Apurva Gandhi, Kyunghyun Cho et al.

Data mixing decides how to combine different sources or types of data and is a consequential problem throughout language model training. In pretraining, data composition is a key determinant of model quality; in continual learning and adaptation, it governs what is retained and acquired. Yet existing data mixing methods address only one phase of this lifecycle at a time: some require smaller proxy models tied to a single training phase, others assume a fixed domain set, and continual learning lacks principled guidance altogether. We argue that data mixing is fundamentally an online decision making problem -- one that recurs throughout training and demands a single, unified solution. We introduce OP-Mix (On-Policy Mix), a data mixing algorithm that operates across the entire language model training lifecycle. Our main insight is that candidate data mixtures can be cheaply simulated by interpolating between low-rank adapters trained directly on the current model, eliminating separate proxy models and ensuring the search is always grounded in the model's actual learning dynamics. Across pretraining, continual midtraining, and continual instruction tuning, OP-Mix consistently finds near-optimal mixtures while using a fraction of the compute of the baselines. In pretraining, OP-Mix improves upon training without mixing by 6.3% in average perplexity. For continual learning, OP-Mix matches the performance of both retraining and on-policy distillation while using 66% and 95% less overall compute, respectively. OP-Mix suggests a different view of language model training: not a sequence of distinct phases, but a single continuous process of learning from data.

Alex Warstadt, Aaron Mueller, Leshem Choshen et al. · ibm-research

Children can acquire language from less than 100 million words of input. Large language models are far less data-efficient: they typically require 3 or 4 orders of magnitude more data and still do not perform as well as humans on many evaluations. These intensive resource demands limit the ability of researchers to train new models and use existing models as developmentally plausible cognitive models. The BabyLM Challenge is a communal effort in which participants compete to optimize language model training on a fixed data budget. Submissions are compared on various evaluation tasks targeting grammatical ability, downstream task performance, and generalization. Participants can submit to up to three tracks with progressively looser data restrictions. From over 30 submissions, we extract concrete recommendations on how best to train data-efficient language models, and on where future efforts should (and perhaps should not) focus. The winning submissions using the LTG-BERT architecture (Samuel et al., 2023) outperformed models trained on trillions of words. Other submissions achieved strong results through training on shorter input sequences or training a student model on a pretrained teacher. Curriculum learning attempts, which accounted for a large number of submissions, were largely unsuccessful, though some showed modest improvements.

Nouha Dziri, Hannah Rashkin, Tal Linzen et al.

Knowledge-grounded dialogue systems powered by large language models often generate responses that, while fluent, are not attributable to a relevant source of information. Progress towards models that do not exhibit this issue requires evaluation metrics that can quantify its prevalence. To this end, we introduce the Benchmark for Evaluation of Grounded INteraction (BEGIN), comprised of 12k dialogue turns generated by neural dialogue systems trained on three knowledge-grounded dialogue corpora. We collect human annotations assessing the extent to which the models' responses can be attributed to the given background information. We then use BEGIN to analyze eight evaluation metrics. We find that these metrics rely on spurious correlations, do not reliably distinguish attributable abstractive responses from unattributable ones, and perform substantially worse when the knowledge source is longer. Our findings underscore the need for more sophisticated and robust evaluation metrics for knowledge-grounded dialogue. We make BEGIN publicly available at https://github.com/google/BEGIN-dataset.

Leshem Choshen, Ryan Cotterell, Michael Y. Hu et al. · ibm-research

After last year's successful BabyLM Challenge, the competition will be hosted again in 2024/2025. The overarching goals of the challenge remain the same; however, some of the competition rules will be different. The big changes for this year's competition are as follows: First, we replace the loose track with a paper track, which allows (for example) non-model-based submissions, novel cognitively-inspired benchmarks, or analysis techniques. Second, we are relaxing the rules around pretraining data, and will now allow participants to construct their own datasets provided they stay within the 100M-word or 10M-word budget. Third, we introduce a multimodal vision-and-language track, and will release a corpus of 50% text-only and 50% image-text multimodal data as a starting point for LM model training. The purpose of this CfP is to provide rules for this year's challenge, explain these rule changes and their rationale in greater detail, give a timeline of this year's competition, and provide answers to frequently asked questions from last year's challenge.

Michael Y. Hu, Aaron Mueller, Candace Ross et al. · ibm-research

The BabyLM Challenge is a community effort to close the data-efficiency gap between human and computational language learners. Participants compete to optimize language model training on a fixed language data budget of 100 million words or less. This year, we released improved text corpora, as well as a vision-and-language corpus to facilitate research into cognitively plausible vision language models. Submissions were compared on evaluation tasks targeting grammatical ability, (visual) question answering, pragmatic abilities, and grounding, among other abilities. Participants could submit to a 10M-word text-only track, a 100M-word text-only track, and/or a 100M-word and image multimodal track. From 31 submissions employing diverse methods, a hybrid causal-masked language model architecture outperformed other approaches. No submissions outperformed the baselines in the multimodal track. In follow-up analyses, we found a strong relationship between training FLOPs and average performance across tasks, and that the best-performing submissions proposed changes to the training data, training objective, and model architecture. This year's BabyLM Challenge shows that there is still significant room for innovation in this setting, in particular for image-text modeling, but community-driven research can yield actionable insights about effective strategies for small-scale language modeling.

Lucas Charpentier, Leshem Choshen, Ryan Cotterell et al. · ibm-research

BabyLM aims to dissolve the boundaries between cognitive modeling and language modeling. We call for both workshop papers and for researchers to join the 3rd BabyLM competition. As in previous years, we call for participants in the data-efficient pretraining challenge in the general track. This year, we also offer a new track: INTERACTION. This new track encourages interactive behavior, learning from a teacher, and adapting the teaching material to the student. We also call for papers outside the competition in any relevant areas. These include training efficiency, cognitively plausible research, weak model evaluation, and more.

William Merrill, Zhaofeng Wu, Norihito Naka et al. · mit

Do LMs infer the semantics of text from co-occurrence patterns in their training data? Merrill et al. (2022) argue that, in theory, sentence co-occurrence probabilities predicted by an optimal LM should reflect the entailment relationship of the constituent sentences, but it is unclear whether probabilities predicted by neural LMs encode entailment in this way because of strong assumptions made by Merrill et al. (namely, that humans always avoid redundancy). In this work, we investigate whether their theory can be used to decode entailment relations from neural LMs. We find that a test similar to theirs can decode entailment relations between natural sentences, well above random chance, though not perfectly, across many datasets and LMs. This suggests LMs implicitly model aspects of semantics to predict semantic effects on sentence co-occurrence patterns. However, we find the test that predicts entailment in practice works in the opposite direction to the theoretical test. We thus revisit the assumptions underlying the original test, finding its derivation did not adequately account for redundancy in human-written text. We argue that better accounting for redundancy related to explanations might derive the observed flipped test and, more generally, improve computational models of speakers in linguistics.

Linlu Qiu, Fei Sha, Kelsey Allen et al.

Artificial intelligence systems based on large language models (LLMs) are increasingly used as agents that interact with users and with the world. To do so successfully, LLMs need to construct internal representations of the world and form probabilistic beliefs about those representations. To provide a user with personalized recommendations, for example, the LLM needs to gradually infer the user's preferences, over the course of multiple interactions. To evaluate whether contemporary LLMs are able to do so, we use the Bayesian inference framework from probability theory, which lays out the optimal way to update an agent's beliefs as it receives new information. We first show that LLMs do not update their beliefs as expected from the Bayesian framework, and that consequently their predictions do not improve as expected as more information becomes available. To address this issue, we teach the LLMs to reason in a Bayesian manner by training them to mimic the predictions of the normative Bayesian model. We find that this approach not only significantly improves the LLM's performance on the particular recommendation task it is trained on, but also enables generalization to other tasks. This suggests that this method teaches the LLM to better approximate Bayesian reasoning. More generally, our results indicate that LLMs can effectively learn reasoning skills from examples and generalize those skills to new domains.

Michael Y. Hu, Jackson Petty, Chuan Shi et al.

Pretraining language models on formal language can improve their acquisition of natural language. Which features of the formal language impart an inductive bias that leads to effective transfer? Drawing on insights from linguistics and complexity theory, we hypothesize that effective transfer occurs when two conditions are met: the formal language should capture the dependency structures present in natural language, and it should remain within the computational limitations of the model architecture. We experiment with pre-pretraining (training on formal language before natural languages) on transformers and find that formal languages capturing hierarchical dependencies indeed enable language models to achieve lower loss on natural language and better linguistic generalization compared to other formal languages. We also find modest support for the hypothesis that the formal language should fall within the computational limitations of the architecture. Strikingly, pre-pretraining reduces loss more efficiently than training on a matched amount of natural language. For a 1B-parameter language model trained on roughly 1.6B tokens of natural language, pre-pretraining achieves the same loss and better linguistic generalization with a 33% smaller token budget. Finally, we also give mechanistic evidence of transfer from formal to natural language: attention heads acquired during pre-pretraining remain crucial for the model's performance on syntactic evaluations.

Lindia Tjuatja, Graham Neubig, Tal Linzen et al.

When comparing the linguistic capabilities of language models (LMs) with humans using LM probabilities, factors such as the length of the sequence and the unigram frequency of lexical items have a significant effect on LM probabilities in ways that humans are largely robust to. Prior works in comparing LM and human acceptability judgments treat these effects uniformly across models, making a strong assumption that models require the same degree of adjustment to control for length and unigram frequency effects. We propose MORCELA, a new linking theory between LM scores and acceptability judgments where the optimal level of adjustment for these effects is estimated from data via learned parameters for length and unigram frequency. We first show that MORCELA outperforms a commonly used linking theory for acceptability - SLOR (Pauls and Klein, 2012; Lau et al. 2017) - across two families of transformer LMs (Pythia and OPT). Furthermore, we demonstrate that the assumed degrees of adjustment in SLOR for length and unigram frequency overcorrect for these confounds, and that larger models require a lower relative degree of adjustment for unigram frequency, though a significant amount of adjustment is still necessary for all models. Finally, our subsequent analysis shows that larger LMs' lower susceptibility to frequency effects can be explained by an ability to better predict rarer words in context.

Qihan Wang, Shidong Pan, Tal Linzen et al.

Large Language Models (LLMs) are known to lack cultural representation and overall diversity in their generations, from expressing opinions to answering factual questions. To mitigate this problem, we propose multilingual prompting: a prompting method which generates several variations of a base prompt with added cultural and linguistic cues from several cultures, generates responses, and then combines the results. Building on evidence that LLMs have language-specific knowledge, multilingual prompting seeks to increase diversity by activating a broader range of cultural knowledge embedded in model training data. Through experiments across multiple models (GPT-4o, GPT-4o-mini, LLaMA 70B, and LLaMA 8B), we show that multilingual prompting consistently outperforms existing diversity-enhancing techniques such as high-temperature sampling, step-by-step recall, and persona prompting. Further analyses show that the benefits of multilingual prompting vary between high and low resource languages and across model sizes, and that aligning the prompting language with cultural cues reduces hallucination about culturally-specific information.

Grusha Prasad, Tal Linzen

Structural priming is a widely used psycholinguistic paradigm to study human sentence representations. In this work we introduce SPAWN, a cognitively motivated parser that can generate quantitative priming predictions from contemporary theories in syntax which assume a lexicalized grammar. By generating and testing priming predictions from competing theoretical accounts, we can infer which assumptions from syntactic theory are useful for characterizing the representations humans build when processing sentences. As a case study, we use SPAWN to generate priming predictions from two theories (Whiz-Deletion and Participial-Phase) which make different assumptions about the structure of English relative clauses. By modulating the reanalysis mechanism that the parser uses and strength of the parser's prior knowledge, we generated nine sets of predictions from each of the two theories. Then, we tested these predictions using a novel web-based comprehension-to-production priming paradigm. We found that while the some of the predictions from the Participial-Phase theory aligned with human behavior, none of the predictions from the the Whiz-Deletion theory did, thus suggesting that the Participial-Phase theory might better characterize human relative clause representations.

Carter Blum, Katja Filippova, Ann Yuan et al. · deepmind

Large language models (LLMs) struggle with cross-lingual knowledge transfer: they hallucinate when asked in one language about facts expressed in a different language during training. This work introduces a controlled setting to study the causes and dynamics of this phenomenon by training small Transformer models from scratch on synthetic multilingual datasets. We identify a learning phase wherein a model develops either separate or unified representations of the same facts across languages, and show that unification is essential for cross-lingual transfer. We also show that the degree of unification depends on mutual information between facts and training data language, and on how easy it is to extract that language. Based on these insights, we develop methods to modulate the level of cross-lingual transfer by manipulating data distribution and tokenization, and we introduce metrics and visualizations to formally characterize their effects on unification. Our work shows how controlled settings can shed light on pre-training dynamics and suggests new directions for improving cross-lingual transfer in LLMs.

Wentao Wang, Guangyuan Jiang, Tal Linzen et al.

Humans can quickly learn a new word from a few illustrative examples, and then systematically and flexibly use it in novel contexts. Yet the abilities of current language models for few-shot word learning, and methods for improving these abilities, are underexplored. In this study, we introduce a novel method, Meta-training for IN-context learNing Of Words (Minnow). This method trains language models to generate new examples of a word's usage given a few in-context examples, using a special placeholder token to represent the new word. This training is repeated on many new words to develop a general word-learning ability. We find that training models from scratch with Minnow on human-scale child-directed language enables strong few-shot word learning, comparable to a large language model (LLM) pre-trained on orders of magnitude more data. Furthermore, through discriminative and generative evaluations, we demonstrate that finetuning pre-trained LLMs with Minnow improves their ability to discriminate between new words, identify syntactic categories of new words, and generate reasonable new usages and definitions for new words, based on one or a few in-context examples. These findings highlight the data efficiency of Minnow and its potential to improve language model performance in word learning tasks.

Jackson Petty, Jaulie Goe, Tal Linzen

Low-resource languages pose a challenge for machine translation with large language models (LLMs), which require large amounts of training data. One potential way to circumvent this data dependence is to rely on LLMs' ability to use in-context descriptions of languages, like textbooks and dictionaries. To do so, LLMs must be able to infer the link between the languages' grammatical descriptions and the sentences in question. Here we isolate this skill using a formal analogue of the task: string transduction based on a formal grammar provided in-context. We construct synchronous context-free grammars which define pairs of formal languages designed to model particular aspects of natural language grammar, morphology, and written representation. Using these grammars, we measure how well LLMs can translate sentences from one formal language into another when given both the grammar and the source-language sentence. We vary the size of the grammar, the lengths of the sentences, the syntactic and morphological properties of the languages, and their written script. We note three key findings. First, LLMs' translation accuracy decreases markedly as a function of grammar size and sentence length. Second, differences in morphology and written representation between the source and target languages can strongly diminish model performance. Third, we examine the types of errors committed by models and find they are most prone to recall the wrong words from the target language vocabulary, hallucinate new words, or leave source-language words untranslated.

Shauli Ravfogel, Gilad Yehudai, Tal Linzen et al.

Recent probing studies reveal that large language models exhibit linear subspaces that separate true from false statements, yet the mechanism behind their emergence is unclear. We introduce a transparent, one-layer transformer toy model that reproduces such truth subspaces end-to-end and exposes one concrete route by which they can arise. We study one simple setting in which truth encoding can emerge: a data distribution where factual statements co-occur with other factual statements (and vice-versa), encouraging the model to learn this distinction in order to lower the LM loss on future tokens. We corroborate this pattern with experiments in pretrained language models. Finally, in the toy setting we observe a two-phase learning dynamic: networks first memorize individual factual associations in a few steps, then -- over a longer horizon -- learn to linearly separate true from false, which in turn lowers language-modeling loss. Together, these results provide both a mechanistic demonstration and an empirical motivation for how and why linear truth representations can emerge in language models.

Jackson Petty, Michael Y. Hu, Wentao Wang et al.

Large language models (LLMs) are increasingly expected to perform tasks based only on a specification of the task provided in context, without examples of inputs and outputs; this ability is referred to as instruction following. We introduce the Recognition of Languages In-Context (RELIC) framework to evaluate instruction following using language recognition: the task of determining if a string is generated by formal grammar. Unlike many standard evaluations of LLMs' ability to use their context, this task requires composing together a large number of instructions (grammar productions) retrieved from the context. Because the languages are synthetic, the task can be increased in complexity as LLMs' skills improve, and new instances can be automatically generated, mitigating data contamination. We evaluate state-of-the-art LLMs on RELIC and find that their accuracy can be reliably predicted from the complexity of the grammar and the individual example strings, and that even the most advanced LLMs currently available show near-chance performance on more complex grammars and samples, in line with theoretical expectations. We also use RELIC to diagnose how LLMs attempt to solve increasingly difficult reasoning tasks, finding that as the complexity of the language recognition task increases, models switch to relying on shallow heuristics instead of following complex instructions.

Byung-Doh Oh, Tal Linzen

When people listen to or read a sentence, they actively make predictions about upcoming words: words that are less predictable are generally read more slowly than predictable ones. The success of large language models (LLMs), which, like humans, make predictions about upcoming words, has motivated exploring the use of these models as cognitive models of human linguistic prediction. Surprisingly, in the last few years, as language models have become better at predicting the next word, their ability to predict human reading behavior has declined. This is because LLMs are able to predict upcoming words much better than people can, leading them to predict lower processing difficulty in reading than observed in human experiments; in other words, mainstream LLMs are 'superhuman' as models of language comprehension. In this position paper, we argue that LLMs' superhumanness is primarily driven by two factors: compared to humans, LLMs have much stronger long-term memory for facts and training examples, and they have much better short-term memory for previous words in the text. We advocate for creating models that have human-like long-term and short-term memory, and outline some possible directions for achieving this goal. Finally, we argue that currently available human data is insufficient to measure progress towards this goal, and outline human experiments that can address this gap.

Aaron Mueller, Tal Linzen

Accurate syntactic representations are essential for robust generalization in natural language. Recent work has found that pre-training can teach language models to rely on hierarchical syntactic features - as opposed to incorrect linear features - when performing tasks after fine-tuning. We test what aspects of pre-training are important for endowing encoder-decoder Transformers with an inductive bias that favors hierarchical syntactic generalizations. We focus on architectural features (depth, width, and number of parameters), as well as the genre and size of the pre-training corpus, diagnosing inductive biases using two syntactic transformation tasks: question formation and passivization, both in English. We find that the number of parameters alone does not explain hierarchical generalization: model depth plays greater role than model width. We also find that pre-training on simpler language, such as child-directed speech, induces a hierarchical bias using an order-of-magnitude less data than pre-training on more typical datasets based on web text or Wikipedia; this suggests that in cognitively plausible language acquisition settings, neural language models may be more data-efficient than previously thought.

Linlu Qiu, Peter Shaw, Panupong Pasupat et al.

Generic unstructured neural networks have been shown to struggle on out-of-distribution compositional generalization. Compositional data augmentation via example recombination has transferred some prior knowledge about compositionality to such black-box neural models for several semantic parsing tasks, but this often required task-specific engineering or provided limited gains. We present a more powerful data recombination method using a model called Compositional Structure Learner (CSL). CSL is a generative model with a quasi-synchronous context-free grammar backbone, which we induce from the training data. We sample recombined examples from CSL and add them to the fine-tuning data of a pre-trained sequence-to-sequence model (T5). This procedure effectively transfers most of CSL's compositional bias to T5 for diagnostic tasks, and results in a model even stronger than a T5-CSL ensemble on two real world compositional generalization tasks. This results in new state-of-the-art performance for these challenging semantic parsing tasks requiring generalization to both natural language variation and novel compositions of elements.

R. Thomas McCoy, Paul Smolensky, Tal Linzen et al.

Current language models can generate high-quality text. Are they simply copying text they have seen before, or have they learned generalizable linguistic abstractions? To tease apart these possibilities, we introduce RAVEN, a suite of analyses for assessing the novelty of generated text, focusing on sequential structure (n-grams) and syntactic structure. We apply these analyses to four neural language models (an LSTM, a Transformer, Transformer-XL, and GPT-2). For local structure - e.g., individual dependencies - model-generated text is substantially less novel than our baseline of human-generated text from each model's test set. For larger-scale structure - e.g., overall sentence structure - model-generated text is as novel or even more novel than the human-generated baseline, but models still sometimes copy substantially, in some cases duplicating passages over 1,000 words long from the training set. We also perform extensive manual analysis showing that GPT-2's novel text is usually well-formed morphologically and syntactically but has reasonably frequent semantic issues (e.g., being self-contradictory).

Wang Zhu, Peter Shaw, Tal Linzen et al.

Neural network models often generalize poorly to mismatched domains or distributions. In NLP, this issue arises in particular when models are expected to generalize compositionally, that is, to novel combinations of familiar words and constructions. We investigate learning representations that facilitate transfer learning from one compositional task to another: the representation and the task-specific layers of the models are strategically trained differently on a pre-finetuning task such that they generalize well on mismatched splits that require compositionality. We apply this method to semantic parsing, using three very different datasets, COGS, GeoQuery and SCAN, used alternately as the pre-finetuning and target task. Our method significantly improves compositional generalization over baselines on the test set of the target task, which is held out during fine-tuning. Ablation studies characterize the utility of the major steps in the proposed algorithm and support our hypothesis.

Laura Aina, Tal Linzen

Temporary syntactic ambiguities arise when the beginning of a sentence is compatible with multiple syntactic analyses. We inspect to which extent neural language models (LMs) exhibit uncertainty over such analyses when processing temporarily ambiguous inputs, and how that uncertainty is modulated by disambiguating cues. We probe the LM's expectations by generating from it: we use stochastic decoding to derive a set of sentence completions, and estimate the probability that the LM assigns to each interpretation based on the distribution of parses across completions. Unlike scoring-based methods for targeted syntactic evaluation, this technique makes it possible to explore completions that are not hypothesized in advance by the researcher. We apply this method to study the behavior of two LMs (GPT2 and an LSTM) on three types of temporary ambiguity, using materials from human sentence processing experiments. We find that LMs can track multiple analyses simultaneously; the degree of uncertainty varies across constructions and contexts. As a response to disambiguating cues, the LMs often select the correct interpretation, but occasional errors point to potential areas of improvement.

Jason Wei, Dan Garrette, Tal Linzen et al.

Pre-trained language models perform well on a variety of linguistic tasks that require symbolic reasoning, raising the question of whether such models implicitly represent abstract symbols and rules. We investigate this question using the case study of BERT's performance on English subject-verb agreement. Unlike prior work, we train multiple instances of BERT from scratch, allowing us to perform a series of controlled interventions at pre-training time. We show that BERT often generalizes well to subject-verb pairs that never occurred in training, suggesting a degree of rule-governed behavior. We also find, however, that performance is heavily influenced by word frequency, with experiments showing that both the absolute frequency of a verb form, as well as the frequency relative to the alternate inflection, are causally implicated in the predictions BERT makes at inference time. Closer analysis of these frequency effects reveals that BERT's behavior is consistent with a system that correctly applies the SVA rule in general but struggles to overcome strong training priors and to estimate agreement features (singular vs. plural) on infrequent lexical items.

Alicia Parrish, Sebastian Schuster, Alex Warstadt et al.

Understanding language requires grasping not only the overtly stated content, but also making inferences about things that were left unsaid. These inferences include presuppositions, a phenomenon by which a listener learns about new information through reasoning about what a speaker takes as given. Presuppositions require complex understanding of the lexical and syntactic properties that trigger them as well as the broader conversational context. In this work, we introduce the Naturally-Occurring Presuppositions in English (NOPE) Corpus to investigate the context-sensitivity of 10 different types of presupposition triggers and to evaluate machine learning models' ability to predict human inferences. We find that most of the triggers we investigate exhibit moderate variability. We further find that transformer-based models draw correct inferences in simple cases involving presuppositions, but they fail to capture the minority of exceptional cases in which human judgments reveal complex interactions between context and triggers.