Muult

Muult is a hypothetical metric used in theoretical studies of multilingual learning to quantify cross-lingual transfer in neural models. Defined as the average normalized mutual information between input tokens and model representations across a set of languages in a shared embedding space, muult aims to capture how well a single model preserves linguistic information across languages.

Etymology: The term muult derives from a blend of multilingual and unified learning transfer. It was proposed in speculative literature to name a cross-lingual transfer quantity that reflects information preservation rather than task-specific accuracy alone.

Properties: Muult is dimensionless and typically bounded between 0 and 1. Higher values indicate stronger cross-lingual information preservation. Values can be sensitive to the choice of languages, data distributions, and representation methods, and may vary with model size and training regime.

Calculation: To compute muult, researchers evaluate a multilingual model on several languages, estimate the mutual information between language-agnostic representations and language-specific tokens, normalize, and then average across languages. Different normalization schemes can yield comparable yet distinct muult estimates.

Applications: In theory, muult provides a complementary signal to accuracy by focusing on information flow across languages. It is used to compare multilingual architectures, diagnose transfer failures, and guide ablation studies in cross-lingual learning.

Limitations: As a synthetic metric, muult depends on modeling choices and data. It is not a universal measure of linguistic competence and should be interpreted within the context of the task, data selection, and evaluation framework. See also mutual information, multilingual NLP, and transfer learning.