ZICQ
中 Log in / Sign up
Newsroom Research & Papers #CRAFTER #Model Correction #Feature Project #LLMs & Foundation Models #Disabled learning

CRAFTER: Mining Residuals of Frozen Forecasters Doubles Corrector Gains

Avatar of Mr.Xu

By Mr.Xu Compiled & Reviewed by Editorial

Published: · 10 views

中文阅读 (Chinese) English Version
Co-reading Select any sentence in the article to highlight, annotate, or copy a quote card. Join discussion ↓

Summary:A new arXiv paper introduces CRAFTER (Corrective Residual Agent with Feature-based Temporal Exploration and Reasoning), a framework for discovering corrective features from the residuals of frozen forecasters. By combining compositional search and LLM-generated features, CRAFTER outperforms dedicated feature engineering systems across six datasets and six backbones, roughly doubling the improvement of the corrector alone and reducing error of the weakest backbones by up to 27%. This work offers


Overview

A new arXiv paper introduces CRAFTER (Corrective Residual Agent with Feature-based Temporal Exploration and Reasoning), a framework for discovering corrective features from the residuals of frozen forecasters. It drives a lightweight post-hoc corrector without expensive fine-tuning.

Core Method

CRAFTER employs two complementary feature generators:

  • Compositional Search: Searches over raw input channels to find structured features.
  • LLM Generation: Uses large language models to propose named feature combinations, binary flags, and executable code snippets.

All candidates are accepted or rejected by a unified validation-grounded gate, and a validation-selected corrector applies the accepted features. The source-agnostic pipeline enables fair evaluation of prior feature engineering systems.

Experimental Results

Across six public datasets and six frozen backbones, CRAFTER surpasses all dedicated feature engineering systems at every feature budget:

  • Roughly doubles the improvement of the corrector alone.
  • Reduces error of the weakest backbones by up to 27%.
  • Gains are robust across different LLM backends and persist on fine-tuned backbones.

Technical Highlights

  • Residual Modeling: Unlike traditional feature engineering that models the data-generating process, CRAFTER models the model-failure process.
  • Source-Agnostic Pipeline: Allows prior systems to be evaluated under identical conditions, enabling attribution of improvements to feature source.
  • Lightweight Correction: Keeps backbone frozen, requiring only a lightweight corrector, suitable for resource-constrained scenarios.

Industry Impact and Developer Advice

For developers with large pretrained models that are difficult to fine-tune, CRAFTER offers an efficient correction solution. Its source-agnostic design also makes it a benchmark tool for evaluating feature engineering methods. Developers are advised to explore its application in time-series forecasting, anomaly detection, and other domains.

Source

Based on arXiv paper 2608.05207 (released 2026-08-07).


Source: ArXiv Machine Learning (cs.LG) (2026-08-07)

— END —

Tags: #CRAFTER #Model Correction #Feature Project #LLMs & Foundation Models #Disabled learning

Editorial & Fact-Checking Note: This article is compiled from primary research, official release documentation, and source papers by the ZICQ Newsroom pipeline with automated entity verification and human editorial review. If you notice any technical inaccuracy, please submit a correction via our corrections policy or email our editorial desk directly.

What's your take on this story?
🚀 Explore & Submit Works

Community Comments

Loading live comments and annotations…