Large Language Model Teaches Visual Students: Cross-Modality Transfer of Fine-Grained Conceptual Knowledge

Researchers have introduced LaViD, a framework that lets a large language model teach a vision model fine-grained concepts without ever seeing an image, according to a preprint posted to arXiv on June 25, 2026 [1][2]. The method, called Language-to-Visual Knowledge Distillation, uses a language-only teacher to generate multiple-choice questions that probe semantic distinctions between visual classes [2]. Each class is then mapped to a soft label distribution over those questions, forming a conceptual signature that guides the vision student through an auxiliary distillation loss [2]. The approach does not require paired multimodal data, a departure from conventional distillation pipelines that rely on image-text pairs [2]. In benchmark tests, LaViD consistently outperformed MaKD, a recent method that distills from vision-language models, across multiple fine-grained benchmarks [2]. It also matched or exceeded the performance of state-of-the-art visual distillation techniques such as DKD and MLKD, with additional gains observed when combined with logit standardization [2]. On the Waterbirds dataset, LaViD substantially improved worst-group accuracy, a signal that the distillation process enhances robustness to spurious correlations [2]. The work lands on arXiv, an open-access repository that hosts preprints across physics, computer science, and related fields [7]. As of November 2024, the repository was receiving roughly 24,000 new articles per month [7]. The LaViD preprint appears under the Computer Vision and Pattern Recognition category, and its abstract page includes experimental tools developed through arXivLabs, a community-collaboration framework launched in 2020 [6][1]. arXivLabs allows third-party developers to build features such as bibliographic explorers and code finders that sit directly on article pages, under guidelines that require adherence to openness and user-data privacy [6]. Code for LaViD has been made available on GitHub, and the authors note that the language-only teacher operates without access to image data throughout the distillation process [2].