Teaching multimodal AI new tasks without forgetting old ones

As multimodal large language models learn new tasks sequentially, they tend to overwrite previously acquired knowledge — a problem called catastrophic forgetting. Existing fixes either store past training data (raising privacy concerns) or add architectural overhead that hurts generalization. Octopus sidesteps both issues with History-Free Gradient Orthogonalization (HiFGO), which constrains new gradient updates to be orthogonal to directions that would disturb prior task parameters, using no historical examples. A two-stage fine-tuning strategy separates task adaptation from regularization, balancing plasticity and stability. On the UCIT continual learning benchmark, Octopus outperforms the previous best by 2.14% on average accuracy and 6.82% on last-task accuracy.