Task Arithmetic in AI Explainer
Task arithmetic represents a fine-tuned model as a change from its original base checkpoint. Subtracting the base parameters from the fine-tuned parameters produces a task vector associated with the learned behavior.
Researchers can scale that vector, add it to another compatible base, combine several vectors, or subtract a vector in an attempt to reduce a behavior. The operations resemble arithmetic, but they occur across high-dimensional model parameters.
The method offers a simple way to study and edit relationships among fine-tuned models without retraining on every combined dataset. It is closely related to model merging and depends on models having aligned architectures and origins.
Task vectors are not clean symbolic concepts. Their effects can entangle with unrelated capabilities, depend strongly on scale, and fail when models are misaligned. Claims about adding or removing skills require behavioral, safety, and out-of-distribution evaluation.
Task arithmetic treats the parameter change created by fine-tuning as a vector that can be added, removed, scaled, or combined with other task vectors.
Key terms include task vector, base model, fine-tuned model, parameter difference, scaling, addition, subtraction, alignment and interference.
Subtracting a base checkpoint from its fine-tuned version defines a task vector.
Compatible vectors can be scaled or combined and then applied back to aligned model weights.
Parameter changes can affect many behaviors beyond the task used to create them.
Testing must look for interference, sensitivity to scale, and failures outside benchmark data.
Explore more "Explainers"
Discover additional explainers across politics, science, business, technology, and other fields. Each explainer breaks down a complex idea into clear, everyday language—helping you better understand how major concepts, systems, and debates shape the world around us.