Skill Neologisms: Towards Skill-based Continual Learning

Modern LLMs show mastery over an ever-growing range of skills, as well as the ability to compose them flexibly. However, extending model capabilities to new skills in a scalable manner is an open problem: fine-tuning and parameter-efficient variants risk catastrophic forgetting, while context-based approaches have limited expressiveness and are constrained by the model's effective context. We explore skill neologisms--soft tokens integrated in the model's vocabulary and optimized to improve capabilities over a specific skill--as a way to selectively acquire new skills without weight updates. We first observe that pre-trained LLMs already exhibit tokens associated with procedural knowledge. We then show on a controlled synthetic task that skill neologisms can be learned to improve model capabilities on specific skills while being composable with out-of-distribution skills, and that independently trained skill neologisms can be composed zero-shot. Finally, we validate zero-shot composition of independently learned skill neologisms on the more realistic natural language setting of the Skill-Mix benchmark. These results suggest that skill neologisms may provide a scalable path towards skill-based continual learning.

Large language models acquire broad capabilities during pre-training, but extending them to new specific skills is difficult: modifying model weights to teach a new skill risks degrading existing ones--a phenomenon known as catastrophic forgetting. Humans, by contrast, build on prior knowledge to learn new things without losing what they already know. We take inspiration from this: rather than updating model weights, we teach new skills by learning new words in the model vocabulary, each trained on a curated dataset focusing on the skill of interest. Like a neologism coined to capture an idea no existing word expresses, these new tokens encode procedural knowledge the model can combine freely, both with its existing capabilities and with other learned tokens, all without modifying any model weights. This provides a path to continuously extending what a model can do, one skill at a time, without the risk of forgetting.