SIFT: A Self-Improving Document Classifier That Retrains Without a Human

Researchers have introduced SIFT (Self-Improving, Frozen-gate Training), a document classification service that keeps training itself on real traffic and retrains without human involvement. The preprint was published on arXiv in July 2026.

How the cheap pipeline works

SIFT deliberately serves classification through a cheap CPU pipeline: the SPLADE sparse encoder feeds features into a LightGBM model, and only the minority of low-confidence pages goes to the expensive LLM judge. The system labels the bulk of documents without calling a large language model, which keeps inference costs low.

• Architecture: SPLADE (sparse encoder) → LightGBM head, entirely on CPU
• Only the low-confidence minority of pages is escalated to the LLM judge
• Onboarding a new document type takes a declarative bundle: a label space, anchor phrases, and a glossary for the judge
• Retraining happens once a month, automatically, with no human
• The marginal cost of labeling trends toward zero

Why accuracy improves over time

SIFT's accuracy accumulates because the LLM judge's verdicts are written back into the labeled corpus, so the expensive model continuously teaches the cheap one. As the corpus grows, the share of escalations to the LLM falls, and the corpus itself is filled from production traffic rather than from an expensive labeling project at the start. The classic blocker of enterprise classification, the need to first have people label thousands of documents, goes away: the corpus builds itself, and accuracy grows with use.

How to avoid silent degradation

The main risk of a self-retraining classifier is silent degradation: after retraining itself, the model may quietly get worse. SIFT addresses this with a two-part promote gate. The first gate checks for F1 regression on critical labels; the second is a frozen golden regression set on which the model never trains. Either gate acts as a veto on rolling out a new version.

"This turns 'retrain monthly with no human' from recklessness into

routine," the SIFT preprint on arXiv states.

What this means

SIFT offers an enterprise recipe in which an expensive LLM acts as teacher for a cheap CPU classifier, and two safety gates make it possible to entrust retraining to automation. If the approach can be reproduced, companies will no longer need a large labeling project before launching a classifier: data and accuracy accumulate on their own as the system runs.

Frequently asked questions

What is SIFT?

SIFT (Self-Improving, Frozen-gate Training) is a dynamic document classification service that keeps training itself on production traffic: a cheap SPLADE + LightGBM pipeline labels the bulk of pages, while an LLM judge handles the disputed cases and adds them to the training corpus.

Is manual document labeling required?

No. Instead of an upfront labeling project for a new document type, a declarative bundle is enough: a label space, anchor phrases, and a glossary for the LLM judge. After that, the corpus grows by itself from real traffic.

How does SIFT protect against degradation during automatic retraining?

With two gates: a check for F1 regression on critical labels, and a frozen golden set on which the model never trains. If at least one gate fails, the new version is not rolled out.