On post-training artificial intelligence foundation models

Artificial intelligence foundation models learn broad, general-purpose skills by pre-training on large datasets. Making such a model good at something new, however, usually means either training a fresh model from scratch, which is expensive, or fine-tuning the existing one, which tends to erase what it already knew. This thesis studies post-training: how to give a pre-trained foundation model a new capability efficiently, without degrading the abilities it already has. We develop six post-training methods across vision, language, and image generation. The first lets a model keep learning new tasks in sequence without forgetting earlier ones, by pairing a shared feature extractor with separate task-specific classifiers that cannot interfere with each other. The second makes adaptation from a handful of examples generalize much better, by treating the instructions learned for a vision-language model as a probability distribution rather than one fixed setting. The third gives text-to-image models control over where objects appear, by composing a scene one object at a time and learning this from automatically generated data. The fourth allows a vision-language model to read long descriptions instead of short ones. The fifth teaches a multimodal language model to point to the image regions it is talking about, while leaving its original skills intact. The sixth makes image generation work natively in six languages rather than English alone. One principle runs through all of them: add dedicated capacity or supervision for the new skill, and shield existing knowledge from interference. Post-training is therefore a practical, economical alternative to ever-larger pre-training.