multitask

Definition

Handling several tasks with one model, so that a shared representation can transfer knowledge between them. The upside is sample efficiency and generality; the downside is interference, where tasks compete for capacity and the joint model underperforms specialists. Broad benchmarks like MMLU exist to measure this. Do not confuse it with multimodal, which is about multiple input types rather than multiple objectives.

Explainers using this term

• LLM Evaluation from Scratch — Reading Benchmarks and the Contamination ProblemMeasuring Massive Multitask Language Understanding