SIFT: самообучающийся классификатор документов, который переобучается без человека

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

How the cheap pipeline works

SIFT serves classification with a deliberately cheap CPU pipeline: the SPLADE sparse encoder feeds features into a LightGBM model, and only a minority of low-confidence pages get escalated to the expensive LLM judge. The system labels the bulk of documents without calling a large language model at all — hence the low inference cost.

• 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 is a declarative bundle: label space, anchor phrases, and a judge glossary
• Retraining happens roughly once a month, automatically, with no human involved
• Marginal labeling cost trends toward zero

Why accuracy improves over time

SIFT's accuracy compounds because the LLM judge's verdicts get written back into the labeled corpus — the expensive model continuously teaches the cheap one. As the corpus grows, the share of escalations to the LLM falls, and the corpus itself keeps growing from production traffic rather than from an expensive upfront labeling project. The classic blocker for enterprise classification — needing people to label thousands of documents first — disappears here: 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: the model can quietly get worse after retraining on its own. SIFT closes 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 that the model never trains on. Either one triggers a veto on rolling out a new version.

"This turns 'retraining once a month with no human' from recklessness

into routine," the SIFT preprint on arXiv states.

What this means

SIFT offers an enterprise recipe where an expensive LLM acts as a teacher for a cheap CPU classifier, and two safety gates make it possible to trust retraining to automation. If the approach is reproduced, companies would no longer need a large-scale labeling project before launching a classifier — data and accuracy accumulate on their own through use.

Frequently Asked Questions

What is SIFT?

SIFT (Self-Improving, Frozen-gate Training) is a dynamic document classification service that fine-tunes itself on production traffic: a cheap SPLADE+LightGBM pipeline labels the bulk of pages, while an LLM judge handles disputed cases and feeds 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. From there, the corpus grows on its own from real traffic.

How is SIFT protected from degradation during auto-retraining?

By two gates: an F1 regression check on critical labels, and a frozen "golden" set that the model never trains on. If even one gate fails, the new version does not get rolled out.