A Classifier That Teaches Itself: Self-Improving, Frozen-gate Training (SIFT) for Dynamic Document Classification

arXiv cs.CL Papers
document-classification self-improving frozen-gate-training sift sparse-encoding llm-judge safety-mechanism

Summary

SIFT introduces a self-improving document classifier that uses a cheap SPLADE-LightGBM pipeline and an LLM judge to continuously teach itself, while a frozen-gate safety mechanism prevents silent regression, enabling autonomous retraining without human labeling overhead.

arXiv:2607.18358v1 Announce Type: new Abstract: Document classification is a solved problem in the laboratory and an unsolved one in the enterprise. The blocker is rarely model architecture; it is the labeling project that must precede a model and the institutional fear of letting a model retrain itself once one exists. We present SIFT (Self-Improving, Frozen-gate Training), a dynamic classifier service, which attacks both. SIFT serves classification from a deliberately cheap, CPU-bound pipeline, a SPLADE sparse encoder feeding a LightGBM head, and escalates only the low-confidence minority of pages to an LLM judge. The judge's verdicts are written back into a labeled corpus, so the expensive model continuously teaches the cheap one: the escalation rate falls, the corpus grows from production traffic rather than from an up-front annotation effort, and accuracy compounds with use. Onboarding a new document family requires only a declarative bundle, label space, anchor phrases, and a judge glossary, not a labeling project. The harder problem is safety: an autonomously retraining classifier can silently regress. SIFT resolves this with a two-part promote gate, a critical-label F1 regression check plus a frozen golden regression set the model is never trained on, either of which vetoes promotion. This turns "retrain monthly without a human" from reckless into routine. We describe the architecture, the self-feeding corpus loop, the frozen-gate promotion mechanism, and an illustrative multi-domain deployment, and we discuss the economics of a classifier whose marginal labeling cost trends toward zero.
Original Article
View Cached Full Text

Cached at: 07/22/26, 08:23 AM 

# A Classifier That Teaches Itself: Self-Improving, Frozen-gate Training (SIFT) for Dynamic Document Classification
Source: [https://arxiv.org/abs/2607.18358](https://arxiv.org/abs/2607.18358)
[View PDF](https://arxiv.org/pdf/2607.18358)

> Abstract:Document classification is a solved problem in the laboratory and an unsolved one in the enterprise\. The blocker is rarely model architecture; it is the labeling project that must precede a model and the institutional fear of letting a model retrain itself once one exists\. We present SIFT \(Self\-Improving, Frozen\-gate Training\), a dynamic classifier service, which attacks both\. SIFT serves classification from a deliberately cheap, CPU\-bound pipeline, a SPLADE sparse encoder feeding a LightGBM head, and escalates only the low\-confidence minority of pages to an LLM judge\. The judge's verdicts are written back into a labeled corpus, so the expensive model continuously teaches the cheap one: the escalation rate falls, the corpus grows from production traffic rather than from an up\-front annotation effort, and accuracy compounds with use\. Onboarding a new document family requires only a declarative bundle, label space, anchor phrases, and a judge glossary, not a labeling project\. The harder problem is safety: an autonomously retraining classifier can silently regress\. SIFT resolves this with a two\-part promote gate, a critical\-label F1 regression check plus a frozen golden regression set the model is never trained on, either of which vetoes promotion\. This turns "retrain monthly without a human" from reckless into routine\. We describe the architecture, the self\-feeding corpus loop, the frozen\-gate promotion mechanism, and an illustrative multi\-domain deployment, and we discuss the economics of a classifier whose marginal labeling cost trends toward zero\.

## Submission history

From: Bogdan Raduta \[[view email](https://arxiv.org/show-email/8c675b53/2607.18358)\] **\[v1\]**Mon, 20 Jul 2026 12:38:50 UTC \(92 KB\)

Similar Articles

Self Improvement via Fast Tree-search

arXiv cs.AI

This paper introduces RecursiveSelfImprovement via Fast Tree-search (SIFT), a sample-efficient framework that uses a lightweight tree-search guided by LLM-as-a-judge evaluations to improve coding agents' performance under budget constraints, outperforming existing methods with lower resource costs.

A Frozen Pixel-Space Diffusion Model Can Guide Itself with Its Own Samples

Hugging Face Daily Papers

This paper introduces Synthetic Self-Guidance (SSG), a method that attaches a lightweight prediction head to a frozen pretrained pixel-space diffusion model, using the discrepancy between intermediate and final predictions as self-guidance during sampling. It shows that model-generated samples suffice for training the head, improving FID by over 50% on several variants without classifier-free guidance and enhancing strong baselines with CFG.

SIA: Self Improving AI with Harness & Weight Updates

Hugging Face Daily Papers

A self-improving AI framework that simultaneously updates both model weights and task-specific agent architecture via a language-model feedback agent, achieving significant gains across legal classification, GPU optimization, and biological denoising tasks.

Self-Distillation Enables Continual Learning [pdf]

Hacker News Top

Introduces Self-Distillation Fine-Tuning (SDFT), a method that enables on-policy learning from demonstrations to achieve continual learning without catastrophic forgetting, outperforming supervised fine-tuning.