A new technique promises to fix a stubborn AI blind spot: language models that answer the same question well in English but stumble when asked in another language.
Researchers built HOTFIXR, a system that uses one model to probe a target model's multilingual weak spots, then generates synthetic training data aimed squarely at those gaps. The team tested it across three in-distribution tasks, three out-of-distribution tasks, and four out-of-distribution languages the model was not trained on. On average, HOTFIXR lifted in-distribution performance by 6.2 percent, cut catastrophic forgetting on out-of-distribution tasks by 3.7 percent, and improved out-of-distribution language performance by 7.1 percent. The code is not public yet; the researchers say it will be released if the paper is accepted.
The result matters because the two standard fixes for this problem both come with a catch. Routing every prompt through English boosts accuracy but flattens a model's ability to reason natively in other languages. Training on language-balanced data spreads performance evenly across languages but tends to drag down overall quality. A method that targets a model's specific weak languages, rather than treating multilingual support as an all-or-nothing tradeoff, would matter to any product serving users who do not type in English.
It is worth remembering this is a single study testing itself against its own benchmark, not a technique running in a shipped chatbot, so how it holds up on languages and tasks outside those four samples is still an open question.