AI TrackerFrontier desk · AEST
Back to news
normalaitech

Measure Before You Train: Typed Decision Models vs. a Classical Baseline vs. a Generative LLM for Routing a Personal Multilingual Inbox

arXiv cs.CL · 24 Oct 2026, 03:49 AEST

A study compares various systems for routing personal multilingual inboxes, including typed decision models and generative LLMs. The findings highlight the importance of accurate measurement protocols and the limitations of different models in predicting email actions.

Key facts

Read aloud

A study compares various systems for routing personal multilingual inboxes, including typed decision models and generative L L M s. The findings highlight the importance of accurate measurement protocols and the limitations of different models in predicting email actions. Personal email agents decide, for every message, what it is, whether it is urgent, whether to draft a reply, create a task, or ignore it. We present Foyer, a self-hosted personal inbox agent in daily use, and a measurement protocol that yields ground truth before any training. With ~two hundred labels, logistic regression has the highest accuracy on every field type zero point seven nine, urgency zero point seven four, needs-reply zero point eight three, is-noise zero point nine zero, though it never predicts the two rarest classes. The sender's relationship is recovered from other emails of the same contact with ninety eight point six percent accuracy where the contact has been seen. Accuracy is not the product metric: composed into actions, the most accurate system most often ignores emails needing action.

Sources in this cluster