LLMs, while outperforming humans in a wide range of tasks, can still fail in unanticipated ways. We focus on two pervasive failure modes: (i) hallucinations, where models produce incorrect information about the world, and (ii) the low-resource effect, where the models show impressive performance in high-resource languages like English but the performance degrades significantly in low-resource languages like Bengali. We study the intersection of these issues and ask: do hallucination detectors suffer from the low-resource effect? We conduct experiments on five tasks across three domains (factual recall, STEM, and Humanities). Experiments with four LLMs and three hallucination detectors reveal a curious finding: As expected, the task accuracies in low-resource languages experience large drops (compared to English). However, the drop in detectors’ accuracy is often several times smaller than the drop in task accuracy. Our findings suggest that even in low-resource languages, the internal mechanisms of LLMs might encode signals about their uncertainty. Further, the detectors are robust within language (even for non-English) and in multilingual setups, but not in cross-lingual settings without in-language supervision.

There are serious attempts at improving the mathematical acumen of LLMs in questions posed in English. In India, where a large fraction of the students study in regional languages, there is a need to assess and improve these state-of-the-art LLMs in their reasoning abilities in regional languages as well. As Hindi is a language predominantly used in India, this study proposes a new dataset on mathematical combinatorics problems consisting of a parallel corpus of problems in English and Hindi collected from NCERT textbooks. We evaluate the “raw” single-shot capabilities of these LLMs in solving problems posed in Hindi. Then we apply a chain-of-thought approach to evaluate the improvement in the abilities of the LLMs at solving combinatorics problems posed in Hindi. Our study reveals that while smaller LLMs like LLaMa3-8B shows a significant drop in performance when questions are posed in Hindi, versus questions posed in English, larger LLMs like GPT4-turbo shows excellent capabilities at solving problems posed in Hindi, almost at par its abilities in English. We make two primary inferences from our study: (1) large models like GPT4 can be readily deployed in schools where Hindi is the primary language of study, especially in rural India; (2) there is a need to improve the multilingual capabilities of smaller models. As these smaller open-source models can be deployed on not so expensive GPUs, it is easier for schools to provide these models to the students, and hence, the latter is an important direction for future research.