Researchers have proposed a new method for training vision-language models that can learn across multiple tasks without losing performance. The approach, called TASSO, uses two techniques to preserve the model's ability to recognize objects and scenes: subspace learning and geometry-aware knowledge distillation. This helps reduce 'catastrophic forgetting' and maintains the model's zero-shot capabilities. --- Why it matters: This matters for engineers working on vision-language models because it addresses a major challenge in AI research: how to train models that can adapt to new tasks without losing their ability to perform existing ones. TASSO's approach could lead to more robust and efficient training of such models. Source: https://arxiv.org/abs/2608.21487