Small-model confidence helps routing on some multilingual tasks

When a small model handles multilingual examples, low confidence might help identify errors a larger model can fix. But if confidence stays high on those errors, routing may miss them; if confidence differs across languages, a global ranking may also allocate larger-model access unevenly.

The authors ask how reliably confidence identifies useful escalations across tasks, model families and languages.

The authors test whether a small model’s answer confidence can choose which multilingual examples to send to a larger same-family model.

When a small model handles multilingual examples, can its confidence show which ones a larger same-family model could fix? The authors test whether routing by confidence works across tasks and whether it sends a balanced share of examples from each language.

At a fixed 50% escalation budget, confidence routing beats random selection on reading comprehension and topic classification, but usually not on knowledge and reasoning.

This task-by-task pattern shows that confidence can help select examples for a larger model, but its value is not consistent across tasks.

The results come from offline, paired predictions for three open model families, four languages and multiple-choice tasks. They do not establish how other confidence measures, models, languages or live serving conditions would perform.