A new arXiv preprint, “Efficiency Measured in English: Tokeniser Coverage Bounds Small Language Models on Indic Scripts,” takes aim at a common practice in multilingual AI. Small language models with quantisation and parameter-efficient adaptation are widely offered as a route to multilingual AI under compute constraints, and their efficiency is almost always reported as a single memory figure. The paper argues that this metric hides a critical issue: tokeniser coverage for Indic scripts may be far worse than for English.
The abstract, as released, stops after stating the research question, so the full findings are not yet available. But the framing is clear: a memory-only view of efficiency ignores how well the tokeniser represents non-English scripts. If tokeniser coverage is poor, the model’s effective capacity for Indic languages is bounded, regardless of the reported size.
The authors appear to be pushing back against the assumption that a small, quantised model is automatically a good multilingual model. Instead, they suggest that tokeniser coverage should be part of any efficiency evaluation, particularly for scripts that are linguistically and typographically distant from English.