Search

The most recent generation of Large Language Models owes its success not only to scale, but also a novel step in their training: reinforcement learning from human feedback (RLHF). In this study, we assessed the impact that this training regime has on the fit between model and human behavior in regards to linguistic behavior. We evaluated three versions of OpenAI's GPT-3 davinci – original, instruction-tuned, and RLHF-trained – using psycholinguistic tasks: subject-verb agreement, sentence acceptability, and event knowledge. We then compared their performance to human participants. We found that the RLHF model is significantly more human-like in its answers, including in the errors it commits. Moreover, the uncertainty of the distribution of its output is closely tied with between-subject variation in humans. This suggests that human feedback improves not only the overall quality of LLMs, but also the alignment between their behavior and the linguistic, metalinguistic, and discursive intuitions of humans.

What predicts individual differences in children’s acquisition of consonant production across languages? Considerations of children’s development of early speech production have traditionally emphasized inherent physiological constraints of the vocal apparatus that speakers generally have in common (i.e., articulatory complexity). In contrast, we propose a statistical learning account of phonological development, in which phonological regularities of the ambient language guide children’s learning of those regularities in production. Across four languages (English, Spanish, Japanese, and Korean), we utilized recent meta-analytic dataset of age of consonant acquisition spanning 28 studies. High-density measures of children's ambient language environment from over 8,000 transcripts of speech directed to over 1,000 children were used to assess how well the frequency of consonants in child-directed speech predict the age of consonant acquisition. Our results suggest that both frequency and articulatory complexity are related to age of acquisition, with similar results found for English, Spanish, Japanese, and Korean. Consonants heard frequently by children tended to be incorporated into their production repertoires earlier and consonants heard less frequently are incorporated into production repertoires later in development. We discuss future directions that incorporate a statistical learning pathway towards learning to produce the sound patterns of the ambient language.

Large language models (LLMs) can be reasonably thought of as models of idealized statistical learners. Thus, the extent to which they grasp the grammar of the language they are trained on suggests how much of it can be learned from memorization, abstraction, and generalization of linguistic input. However, the knowledge of LMMs’ grammar has largely been gleaned from examples of their outputs or datasets not designed to assess how native-like its knowledge is. In this study, we probed the knowledge of an LLM, GPT-3, with a graded grammatical acceptability task previously normed on humans. GPT-3’s ratings were correlated with human ratings, even with minimal examples. Moreover, GPT-3’s deviation from the human norms was predicted by the between-subject variation for each item, and these deviations were rarely outside of the range of human ratings. Follow-up analyses tested the extent to which local probabilistic structure drives these judgments using n-gram models.

[This corrects the article DOI: 10.1098/rsos.190194.].

Assessing scientists using exploitable metrics can lead to the degradation of research methods even without any strategic behaviour on the part of individuals, via 'the natural selection of bad science.' Institutional incentives to maximize metrics like publication quantity and impact drive this dynamic. Removing these incentives is necessary, but institutional change is slow. However, recent developments suggest possible solutions with more rapid onsets. These include what we call open science improvements, which can reduce publication bias and improve the efficacy of peer review. In addition, there have been increasing calls for funders to move away from prestige- or innovation-based approaches in favour of lotteries. We investigated whether such changes are likely to improve the reproducibility of science even in the presence of persistent incentives for publication quantity through computational modelling. We found that modified lotteries, which allocate funding randomly among proposals that pass a threshold for methodological rigour, effectively reduce the rate of false discoveries, particularly when paired with open science improvements that increase the publication of negative results and improve the quality of peer review. In the absence of funding that targets rigour, open science improvements can still reduce false discoveries in the published literature but are less likely to improve the overall culture of research practices that underlie those publications.