When New Generators Arrive: Lifelong Machine-Generated Text Attribution via Ridge Feature Transfer

2026-06-06 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Natural Language Processing · Depth: Expert, extended

Summary

RidgeFT is a lightweight analytic update framework designed for lifelong Machine-Generated Text (MGT) attribution, addressing the challenge of continuously incorporating new text generators while preserving recognition of previously seen ones. This framework trains a task-aware encoder on an initial generator set, then freezes it, storing only compact class-wise sufficient statistics. It employs covariance calibration to suppress generator-irrelevant variations, uses fixed random features to enhance representation capacity, and updates new classes via closed-form ridge regression without requiring exemplar replay. Across multi-topic evaluations using P3, P4, and P5 protocols on MGT-Academic and AIGTBench datasets, RidgeFT consistently outperforms baselines. It achieves a 0.886 full-F1, 0.902 old-class F1, and 0.804 new-class F1 under the P5 protocol, improving full-F1 by 0.037 and new-class F1 by 0.107 over the strongest baselines, demonstrating superior data efficiency.

Key takeaway

For AI Scientists or Machine Learning Engineers developing machine-generated text attribution systems, RidgeFT offers a robust and efficient solution for adapting to continuously emerging large language models. You should consider integrating its replay-free analytic update framework to maintain high accuracy on both old and new generators, especially under low-resource conditions. This approach avoids catastrophic forgetting and costly full model retraining, allowing your systems to evolve scalably. Be mindful of its statistical storage requirements, though low-precision and merged statistics can significantly reduce this footprint.

Key insights

Lifelong MGT attribution benefits from frozen encoders and analytic updates, balancing new-class adaptation with old-class retention without replay.

Principles

• Decouple new generator learning from deep representation updates.
• Stable representations enable efficient incremental knowledge absorption.
• Analytic updates via sufficient statistics avoid catastrophic forgetting.

Method

RidgeFT trains an encoder, freezes it, then uses covariance calibration, isotropic random feature lifting, and class-balanced ridge regression with sufficient statistics for replay-free, closed-form incremental updates.

In practice

• Employ bf16 precision for class statistics to reduce storage.
• Use merged bf16 statistics for new-class-only incremental updates.
• Consider per-class bf16 for updates adding samples to old classes.

Topics

• Machine-Generated Text Attribution
• Lifelong Learning
• Continual Learning
• Ridge Regression
• Random Features
• Large Language Models

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• When New Generators Arrive: Lifelong Machine-Generated Text Attribution via Ridge Feature Transfer — Lifelong MGT attribution can be effectively achieved through feature-stable analytic updates without exemplar replay.
• Practical transfer learning for NLP with spaCy and Prodigy — Modern NLP workflows combine pre-training, active learning, and continuous maintenance to accelerate development and improve model longevity.
• LuminaSFT: Generating Synthetic Fine-Tuning Data for Small Language Models — Synthetic data generation and regeneration significantly improve small language model performance, especially with task-specific approaches.
• ECA: Efficient Continual Alignment for Open-Ended Image-to-Text Generation — ECA enables efficient continual alignment in OpenITG by adapting VLM alignment modules to evolving visual data without prior raw data.
• EvoSelect: Data-Efficient LLM Evolution for Targeted Task Adaptation — EvoSelect efficiently adapts LLMs by selecting diverse, task-aligned synthetic data using optimal transport and diversity regularization.
• Efficient Test-Time Finetuning of LLMs via Convex Reconstruction and Gradient Caching — HullFT uses convex reconstruction and gradient caching to make test-time finetuning of LLMs faster and more efficient.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.