Deep neural networks have become remarkably accurate on vision benchmarks, yet they still struggle with generalization. The authors of a new arXiv preprint argue that a key bottleneck is the supervision signal itself: one-hot labels, which simply name a class, carry little information about why an image belongs to that class or how it relates to others.
To address this, the paper proposes using language guidance—natural language descriptions or embeddings—as a complement or alternative to one-hot labels. The study is framed as an empirical investigation into whether such language-based supervision improves a model's ability to generalize beyond its training distribution.
The abstract is truncated, so specific results and experimental details are not yet available from the source. What is clear is the motivation: richer supervision from language could help models learn more transferable features. The study's contribution will be in testing this hypothesis systematically across vision tasks.