Search for: All records

Recent language models generate false but plausible-sounding text with surprising frequency. Such “hallucinations” are an obstacle to the usability of language-based AI systems and can harm people who rely upon their outputs. This work shows that there is an inherent statistical lower-bound on the rate that pretrained language models hallucinate certain types of facts, having nothing to do with the transformer LM architecture or data quality. For “arbitrary” facts whose veracity cannot be determined from the training data, we show that hallucinations must occur at a certain rate for language models that satisfy a statistical calibration condition appropriate for generative language models. Specifically, if the maximum probability of any fact is bounded, we show that the probability of generating a hallucination is close to the fraction of facts that occur exactly once in the training data (a “Good-Turing” estimate), even assuming ideal training data without errors. One conclusion is that models pretrained to be sufficiently good predictors (i.e., calibrated) may require post-training to mitigate hallucinations on the type of arbitrary facts that tend to appear once in the training set. However, our analysis also suggests that there is no statistical reason that pretraining will lead to hallucination on facts that tend to appear more than once in the training data (like references to publications such as articles and books, whose hallucinations have been particularly notable and problematic) or on systematic facts (like arithmetic calculations). Therefore, different architectures and learning algorithms may mitigate these latter types of hallucinations. 

Large language models have recently been able to perform well in a wide variety of circumstances. In this work, we explore the possibility of large language models, specifically GPT-3, to write explanations for middle-school mathematics problems, with the goal of eventually using this process to rapidly generate explanations for the mathematics problems of new curricula as they emerge, shortening the time to integrate new curricula into online learning platforms. To generate explanations, two approaches were taken. The first approach attempted to summarize the salient advice in tutoring chat logs between students and live tutors. The second approach attempted to generate explanations using few-shot learning from explanations written by teachers for similar mathematics problems. After explanations were generated, a survey was used to compare their quality to that of explanations written by teachers. We test our methodology using the GPT-3 language model. Ultimately, the synthetic explanations were unable to outperform teacher written explanations. In the future more powerful large language models may be employed, and GPT-3 may still be effective as a tool to augment teachers’ process for writing explanations, rather than as a tool to replace them. The explanations, survey results, analysis code, and a dataset of tutoring chat logs are all available at https://osf.io/wh5n9/. 

A generative model may generate utter nonsense when it is fit to maximize the likelihood of observed data. This happens due to “model error,” i.e., when the true data generating distribution does not fit within the class of generative models being learned. To address this, we propose a model of active distribution learning using a binary invalidity oracle that identifies some examples as clearly invalid, together with random positive examples sampled from the true distribution. The goal is to maximize the likelihood of the positive examples subject to the constraint of (almost) never generating examples labeled invalid by the oracle. Guarantees are agnostic compared to a class of probability distributions. We first show that proper learning may require exponentially many queries to the invalidity oracle. We then give an improper distribution learning algorithm that uses only polynomially many queries. 

Mobile devices use language models to suggest words and phrases for use in text entry. Traditional language models are based on contextual word frequency in a static corpus of text. However, certain types of phrases, when offered to writers as suggestions, may be systematically chosen more often than their frequency would predict. In this paper, we propose the task of generating suggestions that writers accept, a related but distinct task to making accurate predictions. Although this task is fundamentally interactive, we propose a counterfactual setting that permits offline training and evaluation. We find that even a simple language model can capture text characteristics that improve acceptability.