The paper, posted on arXiv, explores whether multilingual language models store language information in separable structures. If so, switching between languages could be as simple as applying targeted interventions in the model's activation space.
The authors frame this as a testable hypothesis for multilingual in-context learning. Rather than retraining or fine-tuning, such interventions would offer a lightweight way to control the language of a model's output.
The abstract outlines the hypothesis and the intended study but does not include results. The full paper presumably details the experiments and findings.